R. Adams Dudley, MD, MBA

Study Sites

We examined the performance of 209 hospitals (63% of all 334 non‐federal facilities in California) participating in the California Hospital Assessment and Reporting Taskforce (CHART) at the time of the survey. CHART is a voluntary quality reporting initiative that began publicly reporting hospital quality data in January 2006.

Hospital‐level Organizational, Case‐mix, and Quality Data

Hospital organizational characteristics (eg, bed size) were obtained from publicly available discharge and utilization data sets from the California Office of Statewide Health Planning and Development (OSHPD). We also linked hospital‐level patient‐mix data (eg, race) from these OSHPD files.

We obtained quality of care data from CHART for January 2006 through June 2007, the time period corresponding to the survey. Quality metrics included 16 measures collected by the Center for Medicare and Medicaid Services (www.cms.hhs.gov) and extensively used in quality research.1, 4, 13, 15‐17 Rather than define a single measure, we examined multiple process measures, anticipating differential impacts of hospitalists on various processes of care for AMI, CHF, and pneumonia. Measures were further divided among those that are usually measured upon initial presentation to the hospital and those that are measured throughout the entire hospitalization and discharge. This division reflects the division of care in the hospital, where emergency room physicians are likely to have a more critical role for admission processes.

Survey Process

We surveyed all nonfederal, acute care hospitals in California that participated in CHART.2 We first identified contacts at each site via professional society mailing lists. We then sent web‐based surveys to all with available email addresses and a fax/paper survey to the remainder. We surveyed individuals between October 2006 and April 2007 and repeated the process at intervals of 1 to 3 weeks. For remaining nonrespondents, we placed a direct call unless consent to survey had been specifically refused. We contacted the following persons in sequence: (1) hospital executives or administrative leaders; (2) hospital medicine department leaders; (3) admitting emergency room personnel or medical staff officers; and (4) hospital website information. In the case of multiple responses with disagreement, the hospital/hospitalist leader's response was treated as the primary source. At each step, respondents were asked to answer questions only if they had a direct working knowledge of their hospitalist services.

Survey Data

Our key survey question to all respondents included whether the respondents could confirm their hospitals had at least one hospitalist medicine group. Hospital leaders were also asked to participate in a more comprehensive survey of their organizational and clinical characteristics. Within the comprehensive survey, leaders also provided estimates of the percent of general medical patients admitted by hospitalists. This measure, used in prior surveys of hospital leaders,9 was intended to be an easily understood approximation of the intensity of hospitalist utilization in any given hospital. A more rigorous, direct measure was not feasible due to the complexity of obtaining admission data over such a large, diverse set of hospitals.

Process Performance Measures

AMI measures assessed at admission included aspirin and ‐blocker administration within 24 hours of arrival. AMI measures assessed at discharge included aspirin administration, ‐blocker administration, angiotensin converting enzyme inhibitor (ACE‐I) (or angiotensin receptor blocker [ARB]) administration for left ventricular (LV) dysfunction, and smoking cessation counseling. There were no CHF admission measures. CHF discharge measures included assessment of LV function, the use of an ACE‐I or ARB for LV dysfunction, and smoking cessation counseling. Pneumonia admission measures included the drawing of blood cultures prior to the receipt of antibiotics, timely administration of initial antibiotics (<8 hours), and antibiotics consistent with recommendations. Pneumonia discharge measures included pneumococcal vaccination, flu vaccination, and smoking cessation counseling.

For each performance measure, we quantified the percentage of missed quality opportunities, defined as the number of patients who did not receive a care process divided by the number of eligible patients, multiplied by 100. In addition, we calculated composite scores for admission and discharge measures across each condition. We summed the numerators and denominators of individual performance measures to generate a disease‐specific composite numerator and denominator. Both individual and composite scores were produced using methodology outlined by the Center for Medicare & Medicaid Services.18 In order to retain as representative a sample of hospitals as possible, we calculated composite scores for hospitals that had a minimum of 25 observations in at least 2 of the quality indicators that made up each composite score.

Statistical Analysis

We used chi‐square tests, Student t tests, and Mann‐Whitney tests, where appropriate, to compare hospital‐level characteristics of hospitals that utilized hospitalists vs. those that did not. Similar analyses were performed among the subset of hospitals that utilized hospitalists. Among this subgroup of hospitals, we compared hospital‐level characteristics between hospitals that provided information regarding the percent of patients admitted by hospitalists vs. those who did not provide this information.

We used multivariable, generalized linear regression models to assess the relationship between having at least 1 hospitalist group and the percentage of missed quality of care measures. Because percentages were not normally distributed (ie, a majority of hospitals had few missed opportunities, while a minority had many), multivariable models employed log‐link functions with a gamma distribution.19, 20 Coefficients for our key predictor (presence of hospitalists) were transformed back to the original units (percentage of missed quality opportunities) so that a positive coefficient represented a higher number of quality measures missed relative to hospitals without hospitalists. Models were adjusted for factors previously reported to be associated with care quality. Hospital organizational characteristics included the number of beds, teaching status, registered nursing (RN) hours per adjusted patient day, and hospital ownership (for‐profit vs. not‐for‐profit). Hospital patient mix factors included annual percentage of admissions by insurance status (Medicare, Medicaid, other), annual percentage of admissions by race (white vs. nonwhite), annual percentage of do‐not‐resuscitate status at admission, and mean diagnosis‐related group‐based case‐mix index.21 We additionally adjusted for the number of cardiac catheterizations, a measure that moderately correlates with the number of cardiologists and technology utilization.22‐24 In our subset analysis among those hospitals with hospitalists, our key predictor for regression analyses was the percentage of patients admitted by hospitalists. For ease of interpretation, the percentage of patients admitted by hospitalists was centered on the mean across all respondent hospitals, and we report the effect of increasing by 10% the percentage of patients admitted by hospitalists. Models were adjusted for the same hospital organizational characteristics listed above. For those models, a positive coefficient also meant a higher number of measures missed.

For both sets of predictors, we additionally tested for the presence of interactions between the predictors and hospital bed size (both continuous as well as dichotomized at 150 beds) in composite measure performance, given the possibility that any hospitalist effect may be greater among smaller, resource‐limited hospitals. Tests for interaction were performed with the likelihood ratio test. In addition, to minimize any potential bias or loss of power that might result from limiting the analysis to hospitals with complete data, we used the multivariate imputation by chained equations method, as implemented in STATA 9.2 (StataCorp, College Station, TX), to create 10 imputed datasets.25 Imputation of missing values was restricted to confounding variables. Standard methods were then used to combine results over the 10 imputed datasets. We also applied Bonferroni corrections to composite measure tests based on the number of composites generated (n = 5). Thus, for the 5 inpatient composites created, standard definitions of significance (P 0.05) were corrected by dividing composite P values by 5, requiring P 0.01 for significance. The institutional review board of the University of California, San Francisco, approved the study. All analyses were performed using STATA 9.2.

Characteristics of Participating Sites

There were 209 eligible hospitals. All 209 (100%) hospitals provided data about the presence or absence of hospitalists via at least 1 of our survey strategies. The majority of identification of hospitalist utilization was via contact with either hospital or hospitalist leaders, n = 147 (70.3%). Web‐sites informed hospitalist prevalence in only 3 (1.4%) hospitals. There were 8 (3.8%) occurrences of disagreement between sources, all of which had available hospital/hospitalist leader responses. Only 1 (0.5%) hospital did not have the minimum 25 patients eligible for any disease‐specific quality measures during the data reporting period. Collectively, the remaining 208 hospitals accounted for 81% of California's acute care hospital population.

Comparisons of Sites With Hospitalists and Those Without

A total of 170 hospitals (82%) participating in CHART used hospitalists. Hospitals with and without hospitalists differed by a variety of characteristics (Table 1). Sites with hospitalists were larger, less likely to be for‐profit, had more registered nursing hours per day, and performed more cardiac catheterizations.

Relationship Between Hospitalist Group Utilization and the Percentage of Missed Quality Opportunities

Table 2 shows the frequency of missed quality opportunities in sites with hospitalists compared to those without. In general, for both individual and composite measures of quality, multivariable adjustment modestly attenuated the observed differences between the 2 groups of hospitals. We present only the more conservative adjusted estimates.

Compared to hospitals without hospitalists, those with hospitalists did not have any statistically significant differences in the individual and composite admission measures for each of the disease processes. In contrast, there were statistically significant differences between hospitalist and nonhospitalist sites for many individual cardiac processes of care that typically occur after admission from the emergency room (ie, LV function assessment for CHF) or those that occurred at discharge (ie, aspirin and ACE‐I/ARB at discharge for AMI). Similarly, the composite discharge scores for AMI and CHF revealed better overall process measure performance at sites with hospitalists, although the AMI composite did not meet statistical significance. There were no statistically significant differences between groups for the pneumonia process measures assessed at discharge. In addition, for composite measures there were no statistically significant interactions between hospitalist prevalence and bed size, although there was a trend (P = 0.06) for the CHF discharge composite, with a larger effect of hospitalists among smaller hospitals.

Percent of Patients Admitted by Hospitalists

Of the 171 hospitals with hospitalists, 71 (42%) estimated the percent of patients admitted by their hospitalist physicians. Among the respondents, the mean and median percentages of medical patients admitted by hospitalists were 51% (SD = 25%) and 49% (IQR = 30‐70%), respectively. Thirty hospitals were above the sample mean. Compared to nonrespondent sites, respondent hospitals took care of more white patients; otherwise, respondent and nonrespondent hospitals were similar in terms of bed size, location, performance across each measure, and other observable characteristics (Supporting Information, Appendix 1).

Relationship Between the Estimated Percentages of Medical Patients Admitted by Hospitalists and Missed Quality Opportunities

Table 3 displays the change in missed quality measures associated with each additional 10% of patients estimated to be admitted by hospitalists. A higher estimated percentage of patients admitted by hospitalists was associated with statistically significant improvements in quality of care across a majority of individual measures and for all composite discharge measures regardless of condition. For example, every 10% increase in the mean estimated number of patients admitted by hospitalists was associated with a mean of 0.6% (P < 0.001), 0.5% (P = 0.004), and 1.5% (P = 0.006) fewer missed quality opportunities for AMI, CHF, and pneumonia discharge process measures composites, respectively. In addition, for these composite measures, there were no statistically significant interactions between the estimated percentage of patients admitted by hospitalists and bed size (dichotomized at 150 beds), although there was a trend (P = 0.09) for the AMI discharge composite, with a larger effect of hospitalists among smaller hospitals.

In order to test the robustness of our results, we carried out 2 secondary analyses. First, we used multivariable models to generate a propensity score representing the predicted probability of being assigned to a hospital with hospitalists. We then used the propensity score as an additional covariate in subsequent multivariable models. In addition, we performed a complete‐case analysis (including only hospitals with complete data, n = 204) as a check on the sensitivity of our results to missing data. Neither analysis produced results substantially different from those presented.

Study Sites

We examined the performance of 209 hospitals (63% of all 334 non‐federal facilities in California) participating in the California Hospital Assessment and Reporting Taskforce (CHART) at the time of the survey. CHART is a voluntary quality reporting initiative that began publicly reporting hospital quality data in January 2006.

Hospital‐level Organizational, Case‐mix, and Quality Data

Hospital organizational characteristics (eg, bed size) were obtained from publicly available discharge and utilization data sets from the California Office of Statewide Health Planning and Development (OSHPD). We also linked hospital‐level patient‐mix data (eg, race) from these OSHPD files.

We obtained quality of care data from CHART for January 2006 through June 2007, the time period corresponding to the survey. Quality metrics included 16 measures collected by the Center for Medicare and Medicaid Services (www.cms.hhs.gov) and extensively used in quality research.1, 4, 13, 15‐17 Rather than define a single measure, we examined multiple process measures, anticipating differential impacts of hospitalists on various processes of care for AMI, CHF, and pneumonia. Measures were further divided among those that are usually measured upon initial presentation to the hospital and those that are measured throughout the entire hospitalization and discharge. This division reflects the division of care in the hospital, where emergency room physicians are likely to have a more critical role for admission processes.

Survey Process

We surveyed all nonfederal, acute care hospitals in California that participated in CHART.2 We first identified contacts at each site via professional society mailing lists. We then sent web‐based surveys to all with available email addresses and a fax/paper survey to the remainder. We surveyed individuals between October 2006 and April 2007 and repeated the process at intervals of 1 to 3 weeks. For remaining nonrespondents, we placed a direct call unless consent to survey had been specifically refused. We contacted the following persons in sequence: (1) hospital executives or administrative leaders; (2) hospital medicine department leaders; (3) admitting emergency room personnel or medical staff officers; and (4) hospital website information. In the case of multiple responses with disagreement, the hospital/hospitalist leader's response was treated as the primary source. At each step, respondents were asked to answer questions only if they had a direct working knowledge of their hospitalist services.

Survey Data

Our key survey question to all respondents included whether the respondents could confirm their hospitals had at least one hospitalist medicine group. Hospital leaders were also asked to participate in a more comprehensive survey of their organizational and clinical characteristics. Within the comprehensive survey, leaders also provided estimates of the percent of general medical patients admitted by hospitalists. This measure, used in prior surveys of hospital leaders,9 was intended to be an easily understood approximation of the intensity of hospitalist utilization in any given hospital. A more rigorous, direct measure was not feasible due to the complexity of obtaining admission data over such a large, diverse set of hospitals.

Process Performance Measures

AMI measures assessed at admission included aspirin and ‐blocker administration within 24 hours of arrival. AMI measures assessed at discharge included aspirin administration, ‐blocker administration, angiotensin converting enzyme inhibitor (ACE‐I) (or angiotensin receptor blocker [ARB]) administration for left ventricular (LV) dysfunction, and smoking cessation counseling. There were no CHF admission measures. CHF discharge measures included assessment of LV function, the use of an ACE‐I or ARB for LV dysfunction, and smoking cessation counseling. Pneumonia admission measures included the drawing of blood cultures prior to the receipt of antibiotics, timely administration of initial antibiotics (<8 hours), and antibiotics consistent with recommendations. Pneumonia discharge measures included pneumococcal vaccination, flu vaccination, and smoking cessation counseling.

For each performance measure, we quantified the percentage of missed quality opportunities, defined as the number of patients who did not receive a care process divided by the number of eligible patients, multiplied by 100. In addition, we calculated composite scores for admission and discharge measures across each condition. We summed the numerators and denominators of individual performance measures to generate a disease‐specific composite numerator and denominator. Both individual and composite scores were produced using methodology outlined by the Center for Medicare & Medicaid Services.18 In order to retain as representative a sample of hospitals as possible, we calculated composite scores for hospitals that had a minimum of 25 observations in at least 2 of the quality indicators that made up each composite score.

Statistical Analysis

We used chi‐square tests, Student t tests, and Mann‐Whitney tests, where appropriate, to compare hospital‐level characteristics of hospitals that utilized hospitalists vs. those that did not. Similar analyses were performed among the subset of hospitals that utilized hospitalists. Among this subgroup of hospitals, we compared hospital‐level characteristics between hospitals that provided information regarding the percent of patients admitted by hospitalists vs. those who did not provide this information.

We used multivariable, generalized linear regression models to assess the relationship between having at least 1 hospitalist group and the percentage of missed quality of care measures. Because percentages were not normally distributed (ie, a majority of hospitals had few missed opportunities, while a minority had many), multivariable models employed log‐link functions with a gamma distribution.19, 20 Coefficients for our key predictor (presence of hospitalists) were transformed back to the original units (percentage of missed quality opportunities) so that a positive coefficient represented a higher number of quality measures missed relative to hospitals without hospitalists. Models were adjusted for factors previously reported to be associated with care quality. Hospital organizational characteristics included the number of beds, teaching status, registered nursing (RN) hours per adjusted patient day, and hospital ownership (for‐profit vs. not‐for‐profit). Hospital patient mix factors included annual percentage of admissions by insurance status (Medicare, Medicaid, other), annual percentage of admissions by race (white vs. nonwhite), annual percentage of do‐not‐resuscitate status at admission, and mean diagnosis‐related group‐based case‐mix index.21 We additionally adjusted for the number of cardiac catheterizations, a measure that moderately correlates with the number of cardiologists and technology utilization.22‐24 In our subset analysis among those hospitals with hospitalists, our key predictor for regression analyses was the percentage of patients admitted by hospitalists. For ease of interpretation, the percentage of patients admitted by hospitalists was centered on the mean across all respondent hospitals, and we report the effect of increasing by 10% the percentage of patients admitted by hospitalists. Models were adjusted for the same hospital organizational characteristics listed above. For those models, a positive coefficient also meant a higher number of measures missed.

For both sets of predictors, we additionally tested for the presence of interactions between the predictors and hospital bed size (both continuous as well as dichotomized at 150 beds) in composite measure performance, given the possibility that any hospitalist effect may be greater among smaller, resource‐limited hospitals. Tests for interaction were performed with the likelihood ratio test. In addition, to minimize any potential bias or loss of power that might result from limiting the analysis to hospitals with complete data, we used the multivariate imputation by chained equations method, as implemented in STATA 9.2 (StataCorp, College Station, TX), to create 10 imputed datasets.25 Imputation of missing values was restricted to confounding variables. Standard methods were then used to combine results over the 10 imputed datasets. We also applied Bonferroni corrections to composite measure tests based on the number of composites generated (n = 5). Thus, for the 5 inpatient composites created, standard definitions of significance (P 0.05) were corrected by dividing composite P values by 5, requiring P 0.01 for significance. The institutional review board of the University of California, San Francisco, approved the study. All analyses were performed using STATA 9.2.

Characteristics of Participating Sites

There were 209 eligible hospitals. All 209 (100%) hospitals provided data about the presence or absence of hospitalists via at least 1 of our survey strategies. The majority of identification of hospitalist utilization was via contact with either hospital or hospitalist leaders, n = 147 (70.3%). Web‐sites informed hospitalist prevalence in only 3 (1.4%) hospitals. There were 8 (3.8%) occurrences of disagreement between sources, all of which had available hospital/hospitalist leader responses. Only 1 (0.5%) hospital did not have the minimum 25 patients eligible for any disease‐specific quality measures during the data reporting period. Collectively, the remaining 208 hospitals accounted for 81% of California's acute care hospital population.

Comparisons of Sites With Hospitalists and Those Without

A total of 170 hospitals (82%) participating in CHART used hospitalists. Hospitals with and without hospitalists differed by a variety of characteristics (Table 1). Sites with hospitalists were larger, less likely to be for‐profit, had more registered nursing hours per day, and performed more cardiac catheterizations.

Relationship Between Hospitalist Group Utilization and the Percentage of Missed Quality Opportunities

Table 2 shows the frequency of missed quality opportunities in sites with hospitalists compared to those without. In general, for both individual and composite measures of quality, multivariable adjustment modestly attenuated the observed differences between the 2 groups of hospitals. We present only the more conservative adjusted estimates.

Compared to hospitals without hospitalists, those with hospitalists did not have any statistically significant differences in the individual and composite admission measures for each of the disease processes. In contrast, there were statistically significant differences between hospitalist and nonhospitalist sites for many individual cardiac processes of care that typically occur after admission from the emergency room (ie, LV function assessment for CHF) or those that occurred at discharge (ie, aspirin and ACE‐I/ARB at discharge for AMI). Similarly, the composite discharge scores for AMI and CHF revealed better overall process measure performance at sites with hospitalists, although the AMI composite did not meet statistical significance. There were no statistically significant differences between groups for the pneumonia process measures assessed at discharge. In addition, for composite measures there were no statistically significant interactions between hospitalist prevalence and bed size, although there was a trend (P = 0.06) for the CHF discharge composite, with a larger effect of hospitalists among smaller hospitals.

Percent of Patients Admitted by Hospitalists

Of the 171 hospitals with hospitalists, 71 (42%) estimated the percent of patients admitted by their hospitalist physicians. Among the respondents, the mean and median percentages of medical patients admitted by hospitalists were 51% (SD = 25%) and 49% (IQR = 30‐70%), respectively. Thirty hospitals were above the sample mean. Compared to nonrespondent sites, respondent hospitals took care of more white patients; otherwise, respondent and nonrespondent hospitals were similar in terms of bed size, location, performance across each measure, and other observable characteristics (Supporting Information, Appendix 1).

Relationship Between the Estimated Percentages of Medical Patients Admitted by Hospitalists and Missed Quality Opportunities

Table 3 displays the change in missed quality measures associated with each additional 10% of patients estimated to be admitted by hospitalists. A higher estimated percentage of patients admitted by hospitalists was associated with statistically significant improvements in quality of care across a majority of individual measures and for all composite discharge measures regardless of condition. For example, every 10% increase in the mean estimated number of patients admitted by hospitalists was associated with a mean of 0.6% (P < 0.001), 0.5% (P = 0.004), and 1.5% (P = 0.006) fewer missed quality opportunities for AMI, CHF, and pneumonia discharge process measures composites, respectively. In addition, for these composite measures, there were no statistically significant interactions between the estimated percentage of patients admitted by hospitalists and bed size (dichotomized at 150 beds), although there was a trend (P = 0.09) for the AMI discharge composite, with a larger effect of hospitalists among smaller hospitals.

In order to test the robustness of our results, we carried out 2 secondary analyses. First, we used multivariable models to generate a propensity score representing the predicted probability of being assigned to a hospital with hospitalists. We then used the propensity score as an additional covariate in subsequent multivariable models. In addition, we performed a complete‐case analysis (including only hospitals with complete data, n = 204) as a check on the sensitivity of our results to missing data. Neither analysis produced results substantially different from those presented.

Study Sites

We examined the performance of 209 hospitals (63% of all 334 non‐federal facilities in California) participating in the California Hospital Assessment and Reporting Taskforce (CHART) at the time of the survey. CHART is a voluntary quality reporting initiative that began publicly reporting hospital quality data in January 2006.

Hospital‐level Organizational, Case‐mix, and Quality Data

Hospital organizational characteristics (eg, bed size) were obtained from publicly available discharge and utilization data sets from the California Office of Statewide Health Planning and Development (OSHPD). We also linked hospital‐level patient‐mix data (eg, race) from these OSHPD files.

We obtained quality of care data from CHART for January 2006 through June 2007, the time period corresponding to the survey. Quality metrics included 16 measures collected by the Center for Medicare and Medicaid Services (www.cms.hhs.gov) and extensively used in quality research.1, 4, 13, 15‐17 Rather than define a single measure, we examined multiple process measures, anticipating differential impacts of hospitalists on various processes of care for AMI, CHF, and pneumonia. Measures were further divided among those that are usually measured upon initial presentation to the hospital and those that are measured throughout the entire hospitalization and discharge. This division reflects the division of care in the hospital, where emergency room physicians are likely to have a more critical role for admission processes.

Survey Process

We surveyed all nonfederal, acute care hospitals in California that participated in CHART.2 We first identified contacts at each site via professional society mailing lists. We then sent web‐based surveys to all with available email addresses and a fax/paper survey to the remainder. We surveyed individuals between October 2006 and April 2007 and repeated the process at intervals of 1 to 3 weeks. For remaining nonrespondents, we placed a direct call unless consent to survey had been specifically refused. We contacted the following persons in sequence: (1) hospital executives or administrative leaders; (2) hospital medicine department leaders; (3) admitting emergency room personnel or medical staff officers; and (4) hospital website information. In the case of multiple responses with disagreement, the hospital/hospitalist leader's response was treated as the primary source. At each step, respondents were asked to answer questions only if they had a direct working knowledge of their hospitalist services.

Survey Data

Our key survey question to all respondents included whether the respondents could confirm their hospitals had at least one hospitalist medicine group. Hospital leaders were also asked to participate in a more comprehensive survey of their organizational and clinical characteristics. Within the comprehensive survey, leaders also provided estimates of the percent of general medical patients admitted by hospitalists. This measure, used in prior surveys of hospital leaders,9 was intended to be an easily understood approximation of the intensity of hospitalist utilization in any given hospital. A more rigorous, direct measure was not feasible due to the complexity of obtaining admission data over such a large, diverse set of hospitals.

Process Performance Measures

AMI measures assessed at admission included aspirin and ‐blocker administration within 24 hours of arrival. AMI measures assessed at discharge included aspirin administration, ‐blocker administration, angiotensin converting enzyme inhibitor (ACE‐I) (or angiotensin receptor blocker [ARB]) administration for left ventricular (LV) dysfunction, and smoking cessation counseling. There were no CHF admission measures. CHF discharge measures included assessment of LV function, the use of an ACE‐I or ARB for LV dysfunction, and smoking cessation counseling. Pneumonia admission measures included the drawing of blood cultures prior to the receipt of antibiotics, timely administration of initial antibiotics (<8 hours), and antibiotics consistent with recommendations. Pneumonia discharge measures included pneumococcal vaccination, flu vaccination, and smoking cessation counseling.

For each performance measure, we quantified the percentage of missed quality opportunities, defined as the number of patients who did not receive a care process divided by the number of eligible patients, multiplied by 100. In addition, we calculated composite scores for admission and discharge measures across each condition. We summed the numerators and denominators of individual performance measures to generate a disease‐specific composite numerator and denominator. Both individual and composite scores were produced using methodology outlined by the Center for Medicare & Medicaid Services.18 In order to retain as representative a sample of hospitals as possible, we calculated composite scores for hospitals that had a minimum of 25 observations in at least 2 of the quality indicators that made up each composite score.

Statistical Analysis

We used chi‐square tests, Student t tests, and Mann‐Whitney tests, where appropriate, to compare hospital‐level characteristics of hospitals that utilized hospitalists vs. those that did not. Similar analyses were performed among the subset of hospitals that utilized hospitalists. Among this subgroup of hospitals, we compared hospital‐level characteristics between hospitals that provided information regarding the percent of patients admitted by hospitalists vs. those who did not provide this information.

We used multivariable, generalized linear regression models to assess the relationship between having at least 1 hospitalist group and the percentage of missed quality of care measures. Because percentages were not normally distributed (ie, a majority of hospitals had few missed opportunities, while a minority had many), multivariable models employed log‐link functions with a gamma distribution.19, 20 Coefficients for our key predictor (presence of hospitalists) were transformed back to the original units (percentage of missed quality opportunities) so that a positive coefficient represented a higher number of quality measures missed relative to hospitals without hospitalists. Models were adjusted for factors previously reported to be associated with care quality. Hospital organizational characteristics included the number of beds, teaching status, registered nursing (RN) hours per adjusted patient day, and hospital ownership (for‐profit vs. not‐for‐profit). Hospital patient mix factors included annual percentage of admissions by insurance status (Medicare, Medicaid, other), annual percentage of admissions by race (white vs. nonwhite), annual percentage of do‐not‐resuscitate status at admission, and mean diagnosis‐related group‐based case‐mix index.21 We additionally adjusted for the number of cardiac catheterizations, a measure that moderately correlates with the number of cardiologists and technology utilization.22‐24 In our subset analysis among those hospitals with hospitalists, our key predictor for regression analyses was the percentage of patients admitted by hospitalists. For ease of interpretation, the percentage of patients admitted by hospitalists was centered on the mean across all respondent hospitals, and we report the effect of increasing by 10% the percentage of patients admitted by hospitalists. Models were adjusted for the same hospital organizational characteristics listed above. For those models, a positive coefficient also meant a higher number of measures missed.

For both sets of predictors, we additionally tested for the presence of interactions between the predictors and hospital bed size (both continuous as well as dichotomized at 150 beds) in composite measure performance, given the possibility that any hospitalist effect may be greater among smaller, resource‐limited hospitals. Tests for interaction were performed with the likelihood ratio test. In addition, to minimize any potential bias or loss of power that might result from limiting the analysis to hospitals with complete data, we used the multivariate imputation by chained equations method, as implemented in STATA 9.2 (StataCorp, College Station, TX), to create 10 imputed datasets.25 Imputation of missing values was restricted to confounding variables. Standard methods were then used to combine results over the 10 imputed datasets. We also applied Bonferroni corrections to composite measure tests based on the number of composites generated (n = 5). Thus, for the 5 inpatient composites created, standard definitions of significance (P 0.05) were corrected by dividing composite P values by 5, requiring P 0.01 for significance. The institutional review board of the University of California, San Francisco, approved the study. All analyses were performed using STATA 9.2.

Characteristics of Participating Sites

There were 209 eligible hospitals. All 209 (100%) hospitals provided data about the presence or absence of hospitalists via at least 1 of our survey strategies. The majority of identification of hospitalist utilization was via contact with either hospital or hospitalist leaders, n = 147 (70.3%). Web‐sites informed hospitalist prevalence in only 3 (1.4%) hospitals. There were 8 (3.8%) occurrences of disagreement between sources, all of which had available hospital/hospitalist leader responses. Only 1 (0.5%) hospital did not have the minimum 25 patients eligible for any disease‐specific quality measures during the data reporting period. Collectively, the remaining 208 hospitals accounted for 81% of California's acute care hospital population.

Comparisons of Sites With Hospitalists and Those Without

A total of 170 hospitals (82%) participating in CHART used hospitalists. Hospitals with and without hospitalists differed by a variety of characteristics (Table 1). Sites with hospitalists were larger, less likely to be for‐profit, had more registered nursing hours per day, and performed more cardiac catheterizations.

Relationship Between Hospitalist Group Utilization and the Percentage of Missed Quality Opportunities

Table 2 shows the frequency of missed quality opportunities in sites with hospitalists compared to those without. In general, for both individual and composite measures of quality, multivariable adjustment modestly attenuated the observed differences between the 2 groups of hospitals. We present only the more conservative adjusted estimates.

Compared to hospitals without hospitalists, those with hospitalists did not have any statistically significant differences in the individual and composite admission measures for each of the disease processes. In contrast, there were statistically significant differences between hospitalist and nonhospitalist sites for many individual cardiac processes of care that typically occur after admission from the emergency room (ie, LV function assessment for CHF) or those that occurred at discharge (ie, aspirin and ACE‐I/ARB at discharge for AMI). Similarly, the composite discharge scores for AMI and CHF revealed better overall process measure performance at sites with hospitalists, although the AMI composite did not meet statistical significance. There were no statistically significant differences between groups for the pneumonia process measures assessed at discharge. In addition, for composite measures there were no statistically significant interactions between hospitalist prevalence and bed size, although there was a trend (P = 0.06) for the CHF discharge composite, with a larger effect of hospitalists among smaller hospitals.

Percent of Patients Admitted by Hospitalists

Of the 171 hospitals with hospitalists, 71 (42%) estimated the percent of patients admitted by their hospitalist physicians. Among the respondents, the mean and median percentages of medical patients admitted by hospitalists were 51% (SD = 25%) and 49% (IQR = 30‐70%), respectively. Thirty hospitals were above the sample mean. Compared to nonrespondent sites, respondent hospitals took care of more white patients; otherwise, respondent and nonrespondent hospitals were similar in terms of bed size, location, performance across each measure, and other observable characteristics (Supporting Information, Appendix 1).

Relationship Between the Estimated Percentages of Medical Patients Admitted by Hospitalists and Missed Quality Opportunities

Table 3 displays the change in missed quality measures associated with each additional 10% of patients estimated to be admitted by hospitalists. A higher estimated percentage of patients admitted by hospitalists was associated with statistically significant improvements in quality of care across a majority of individual measures and for all composite discharge measures regardless of condition. For example, every 10% increase in the mean estimated number of patients admitted by hospitalists was associated with a mean of 0.6% (P < 0.001), 0.5% (P = 0.004), and 1.5% (P = 0.006) fewer missed quality opportunities for AMI, CHF, and pneumonia discharge process measures composites, respectively. In addition, for these composite measures, there were no statistically significant interactions between the estimated percentage of patients admitted by hospitalists and bed size (dichotomized at 150 beds), although there was a trend (P = 0.09) for the AMI discharge composite, with a larger effect of hospitalists among smaller hospitals.

In order to test the robustness of our results, we carried out 2 secondary analyses. First, we used multivariable models to generate a propensity score representing the predicted probability of being assigned to a hospital with hospitalists. We then used the propensity score as an additional covariate in subsequent multivariable models. In addition, we performed a complete‐case analysis (including only hospitals with complete data, n = 204) as a check on the sensitivity of our results to missing data. Neither analysis produced results substantially different from those presented.