Insurance Status and Hospital Care

Design and Subjects

We examined data from the 2005 Nationwide Inpatient Sample (NIS), a nationally representative database of hospital inpatient stays maintained by the Agency for Healthcare Research and Quality (AHRQ) as part of the Healthcare Cost and Utilization Project (HCUP).20, 21 The NIS is a stratified probability sample of 20% of all US community hospitals, including public hospitals, academic medical centers, and specialty hospitals. Long‐term care hospitals, psychiatric hospitals, and alcoholism/chemical‐dependency treatment facilities are excluded. The 2005 NIS contains data on 7,995,048 discharges from 1054 hospitals located in 37 States and is designed to be representative of all acute care discharges from all US community hospitals.21

We identified discharges with a principal diagnosis of acute myocardial infarction (AMI), stroke, and pneumonia using International Classification of Diseases, 9th Revision, Clinical Modification (ICD‐9‐CM) codes specified in the AHRQ definitions of Inpatient Quality Indicators (Supporting Information Appendix).22 These 3 conditions are among the leading causes of noncancer inpatient deaths in patients under 65 years old,23 and evidence suggests that high mortality may be associated with deficiencies in the quality of inpatient care.24

We confined our analysis to patients 18 to 64 years of age, since this population is most at risk of being uninsured or underinsured.25 We excluded pregnant women because they account for an unusually high proportion of uninsured discharges and were relatively few in our cohort.26 In addition, we excluded patients transferred to another acute care hospital and patients missing payer source and discharge disposition. Our study protocol was approved by the Partners Human Research Committee.

Study Variables

We categorized insurance status as privately insured, uninsured, Medicaid, or Medicare. We defined privately insured patients as those having either Blue Cross or another commercial carrier listed as the primary payer and uninsured patients as those having either no charge or self‐pay listed as the primary payer.27 Other governmental payer categories were noted to share several characteristics with Medicare patients and comprised only a small proportion of the sample, and were thus included with Medicare. In order to account for NIS's sampling scheme and accurately apply sample weights in our analysis, we used Medicare as a separate category. However, since Medicare patients age 18 to 64 years represent a fundamentally different population that is primarily disabled or very ill, only results of privately insured, uninsured, and Medicaid patients are reported.

We selected in‐hospital mortality as the outcome measure and LOS and cost per hospitalization as measures of resource utilization. The NIS includes a binary indicator variable for in‐hospital mortality and specifies inpatient LOS in integers, with same‐day stays coded as 0. NIS's cost estimates are based on hospital cost reports submitted to the Centers for Medicare and Medicaid Services. To test the validity of our cost analyses, we performed parallel analyses using hospital charges as a measure of utilization (charges include hospital overhead, charity care, and bad debt). The resulting adjusted ratios differed little from cost ratios and we opted to report only the details of our cost analyses.

In order to assess the independent association between insurance status and the outcome measures listed above, we selected covariates for inclusion in multivariable models based on the existing literature. Patient covariates included: age group (18‐34 years, 35‐49 years, 50‐64 years), sex (male/female), race/ethnicity (non‐Hispanic white, non‐Hispanic black, Hispanic, other, missing), median income by zip code of residence (less than $37,000, $37,000‐$45,999, $46,000‐$60,999, $61,000 or more), admission through the emergency department (yes/no), admission on a weekend (yes/no), measures of disease severity, and comorbidity indicators. Measures of disease severity specific to each outcome are assigned in the NIS using criteria developed by Medstat (Medstat Disease Staging Software Version 5.2, Thomson Medstat Inc., Ann Arbor, MI). Severity is categorized into 5 levels, with a higher level indicating greater risk. We recorded comorbidities for each patient in our sample using HCUP Comorbidity Software, Version 3.2 (www.hcup‐us.ahrq.gov/tools_software.jsp) based on comorbidity measures reported by Elixhauser et al.28

Hospital covariates included: bed size (small, medium, large), ownership/control (private, government, private or government), geographic region (northeast, midwest, south, west), teaching status (teaching, non‐teaching), and the proportion of uninsured and Medicaid patients (combined) admitted to each hospital for AMI, stroke, or pneumonia. The actual number of hospital beds in each bed size category varies according to a hospital's geographic region and teaching status.27 Ownership/control, geographic region, and teaching status are assigned according to information from the American Hospital Association Annual Survey of Hospitals. The proportion of uninsured and Medicaid patients admitted to each hospital was found to have a nonmonotonic relationship with the outcomes being assessed and was thus treated as a 6‐level categorical variable with the following levels: 0% to 10%, 11% to 20%, 21% to 30%, 31% to 40%, 41% to 50%, and 51% to 100%.

Statistical Analysis

Summary statistics were constructed at the patient level and differences in proportions were evaluated with the chi‐square test. We employed direct standardization, using the age and sex distribution of the entire cohort, to compute age‐standardized and sex‐standardized estimates for each insurance group and compared them using the chi‐square test for in‐hospital mortality and t test for log transformed LOS and cost per hospitalization. For each condition, we developed multivariable logistic regression models for in‐hospital mortality and multivariable linear regression models for log transformed LOS and cost. The patient was the unit of analysis in all models.

In order to elucidate the contribution of disease severity and comorbidities and the proportion of uninsured and Medicaid patients admitted to each hospital, we fitted 3 sequential models for each outcome measure: Model 1 adjusted for patient sociodemographic characteristics and hospital characteristics with the exception of the covariate for the proportion of uninsured and Medicaid patients, Model 2 adjusted for all covariates in the preceding model as well as patients' comorbidities and severity of principal diagnosis, and Model 3 adjusted for all previously mentioned covariates as well as the proportion of uninsured and Medicaid patients admitted to each hospital. We excluded patients who died during hospitalization from the models for LOS and cost. We exponentiated the effect estimates from the log transformed linear regression models so that the adjusted ratio represents the percentage change in the mean LOS or mean cost.

To determine whether the association between patients' insurance status and in‐hospital mortality was modified by seeking care in hospitals treating a smaller or larger proportion of uninsured and Medicaid patients, we entered an interaction term for insurance status and proportion of uninsured and Medicaid patients in the final models (Model 3) for our primary outcome of in‐hospital mortality. However, since no significant interaction was found for any of the 3 conditions, this term was removed from the models and results from the interaction models are not described. In order to assess model specification for the linear regression models, we evaluated the normality of model residuals and found that these were approximately normally distributed. Lastly, we attempted to test the robustness of our analyses by creating fixed effects models that controlled for hospital site but were unable to do so due to the computational limitations of available software packages that could not render fixed effects models with more than 1000 hospital sites.

For all variables except race/ethnicity, data were missing for less than 3% of patients, so we excluded these individuals from adjusted analyses. However, race/ethnicity data were missing for 29% of the sample and were analyzed in 3 different ways, namely, with the missing data treated as a separate covariate level, with the missing data removed from the analysis, and with the missing data assigned to the majority covariate level (white race). The results of our analysis were unchanged no matter how the missing values were assigned. As a result, missing values for race/ethnicity were treated as a separate covariate level in the final analysis.15 Sociodemographic characteristics of patients with missing race/ethnicity information were similar to those with complete data.

We used SUDAAN (Release 9.0.1, Research Triangle Institute, NC) to account for NIS's sampling scheme and generalized estimating equations to adjust for the clustering of patients within hospitals and hospitals within sampling strata.29 In order to account for NIS's stratified probability sampling scheme, SUDAAN uses Taylor series linearization for robust variance estimation of descriptive statistics and regression parameters.30, 31 We present 2‐tailed P values or 95% confidence intervals (CIs) for all statistical comparisons.

Patient and Hospital Characteristics

The final cohort comprised of 154,381 patients discharged from 1018 hospitals in 37 states during calendar year 2005 (Table 1). This cohort was representative of 755,346 working‐age Americans, representing approximately 225,947 cases of AMI (29.9%), 151,812 cases of stroke (20.1%), and 377,588 cases of pneumonia (50.0%). Of these patients, 47.5% were privately insured, 12.0% were uninsured, 17.0% received Medicaid, and 23.5% were assigned to Medicare. Compared with the privately insured, uninsured and Medicaid patients were generally younger, less likely to be white, more likely to have lower income, and more likely to be admitted through the emergency department. Of the 1018 hospitals included in our study, close to half (44.3%) were small, with bedsize ranging from 24 to 249. A large number of hospitals were located in the South (39.9%), and 14.9% were designated teaching hospitals.

Compared with privately insured patients, a larger proportion of uninsured and Medicaid patients had higher predicted mortality levels (Table 2). Medicaid patients had a disproportionately higher predicted LOS, whereas predicted resource demand was higher among privately insured patients. Hypertension (48%), chronic pulmonary disease (29.5%), and uncomplicated diabetes (21.5%) were the 3 most common comorbidities in the study cohort, with a generally higher prevalence of comorbidities among Medicaid patients.

In‐Hospital Mortality

Compared with the privately insured, age‐standardized and sex‐standardized in‐hospital mortality for AMI and stroke was significantly higher for uninsured and Medicaid patients (Table 3). Among pneumonia patients, Medicaid recipients had significantly higher in‐hospital mortality compared with privately insured and uninsured patients.

After multivariable adjustment for additional patient and hospital characteristics, uninsured AMI and stroke patients continued to have significantly higher in‐hospital mortality compared with the privately insured (Table 4). Among pneumonia patients, Medicaid recipients persisted in having significantly higher in‐hospital mortality than the privately insured.

LOS

Among AMI and stroke patients, age‐standardized and sex‐standardized mean LOS was significantly longer for the uninsured and Medicaid recipients compared with the privately insured (Table 3). Among pneumonia patients, the uninsured had a slightly shorter mean LOS compared with the privately insured whereas Medicaid recipients averaged the longest LOS.

These insurance‐related disparities in LOS among pneumonia patients persisted after multivariable adjustment (Table 4). Among AMI patients, only Medicaid recipients persisted in having a significantly longer LOS than the privately insured. Among stroke patients, both the uninsured and Medicaid recipients averaged a longer LOS compared with the privately insured.

Cost per Episode

For all 3 conditions, the uninsured had significantly lower age‐standardized and sex‐standardized costs compared with the privately insured (Table 3). However, Medicaid patients had higher costs than the privately insured for all three conditions, significantly so among patients with stroke and pneumonia.

These insurance‐related disparities in costs persisted in multivariable analyses (Table 4). The uninsured continued to have lower costs compared with the privately insured, significantly so for patients with AMI and pneumonia. Among stroke and pneumonia patients, Medicaid recipients continued to accrue higher costs than the uninsured or privately insured.

Limitations

Our study has several limitations. Since the NIS does not assign unique patient identifiers that would permit tracking of readmissions, we excluded patients transferred to another acute‐care hospital from our study to avoid counting the same patient twice. However, only 10% of hospitalized patients underwent transfer for cardiac procedures in the National Registry of Myocardial Infarction, with privately insured patients more likely to be transferred than other insurance groups.17 Since these patients are also more likely to have better survival, their exclusion likely biased our study toward the null. The same is probable for stroke patients as well.

Some uninsured patients begin Medicaid coverage during hospitalization and should ideally be counted as uninsured but were included under Medicaid in our analysis. They are also likely to be state‐ and plan‐specific variations in Medicaid and private payer coverage that we could not incorporate into our analysis. In addition, we were unable to include deaths that may have occurred shortly after discharge, even though these may have been related to the quality of hospital care. Furthermore, although the 3 conditions we studied are common and responsible for a large number of hospital deaths, they make up about 8% of total annual hospital discharges,23 and caution should be exercised in generalizing our findings to the full spectrum of hospitalizations. Lastly, it is possible that unmeasured confounding could be responsible for the observed associations. Uninsured and Medicaid patients are likely to have more severe disease, which may not be adequately captured by the administrative data available in the NIS. If so, this would explain the mortality association rather than insurance status.36, 37

Design and Subjects

We examined data from the 2005 Nationwide Inpatient Sample (NIS), a nationally representative database of hospital inpatient stays maintained by the Agency for Healthcare Research and Quality (AHRQ) as part of the Healthcare Cost and Utilization Project (HCUP).20, 21 The NIS is a stratified probability sample of 20% of all US community hospitals, including public hospitals, academic medical centers, and specialty hospitals. Long‐term care hospitals, psychiatric hospitals, and alcoholism/chemical‐dependency treatment facilities are excluded. The 2005 NIS contains data on 7,995,048 discharges from 1054 hospitals located in 37 States and is designed to be representative of all acute care discharges from all US community hospitals.21

We identified discharges with a principal diagnosis of acute myocardial infarction (AMI), stroke, and pneumonia using International Classification of Diseases, 9th Revision, Clinical Modification (ICD‐9‐CM) codes specified in the AHRQ definitions of Inpatient Quality Indicators (Supporting Information Appendix).22 These 3 conditions are among the leading causes of noncancer inpatient deaths in patients under 65 years old,23 and evidence suggests that high mortality may be associated with deficiencies in the quality of inpatient care.24

We confined our analysis to patients 18 to 64 years of age, since this population is most at risk of being uninsured or underinsured.25 We excluded pregnant women because they account for an unusually high proportion of uninsured discharges and were relatively few in our cohort.26 In addition, we excluded patients transferred to another acute care hospital and patients missing payer source and discharge disposition. Our study protocol was approved by the Partners Human Research Committee.

Study Variables

We categorized insurance status as privately insured, uninsured, Medicaid, or Medicare. We defined privately insured patients as those having either Blue Cross or another commercial carrier listed as the primary payer and uninsured patients as those having either no charge or self‐pay listed as the primary payer.27 Other governmental payer categories were noted to share several characteristics with Medicare patients and comprised only a small proportion of the sample, and were thus included with Medicare. In order to account for NIS's sampling scheme and accurately apply sample weights in our analysis, we used Medicare as a separate category. However, since Medicare patients age 18 to 64 years represent a fundamentally different population that is primarily disabled or very ill, only results of privately insured, uninsured, and Medicaid patients are reported.

We selected in‐hospital mortality as the outcome measure and LOS and cost per hospitalization as measures of resource utilization. The NIS includes a binary indicator variable for in‐hospital mortality and specifies inpatient LOS in integers, with same‐day stays coded as 0. NIS's cost estimates are based on hospital cost reports submitted to the Centers for Medicare and Medicaid Services. To test the validity of our cost analyses, we performed parallel analyses using hospital charges as a measure of utilization (charges include hospital overhead, charity care, and bad debt). The resulting adjusted ratios differed little from cost ratios and we opted to report only the details of our cost analyses.

In order to assess the independent association between insurance status and the outcome measures listed above, we selected covariates for inclusion in multivariable models based on the existing literature. Patient covariates included: age group (18‐34 years, 35‐49 years, 50‐64 years), sex (male/female), race/ethnicity (non‐Hispanic white, non‐Hispanic black, Hispanic, other, missing), median income by zip code of residence (less than $37,000, $37,000‐$45,999, $46,000‐$60,999, $61,000 or more), admission through the emergency department (yes/no), admission on a weekend (yes/no), measures of disease severity, and comorbidity indicators. Measures of disease severity specific to each outcome are assigned in the NIS using criteria developed by Medstat (Medstat Disease Staging Software Version 5.2, Thomson Medstat Inc., Ann Arbor, MI). Severity is categorized into 5 levels, with a higher level indicating greater risk. We recorded comorbidities for each patient in our sample using HCUP Comorbidity Software, Version 3.2 (www.hcup‐us.ahrq.gov/tools_software.jsp) based on comorbidity measures reported by Elixhauser et al.28

Hospital covariates included: bed size (small, medium, large), ownership/control (private, government, private or government), geographic region (northeast, midwest, south, west), teaching status (teaching, non‐teaching), and the proportion of uninsured and Medicaid patients (combined) admitted to each hospital for AMI, stroke, or pneumonia. The actual number of hospital beds in each bed size category varies according to a hospital's geographic region and teaching status.27 Ownership/control, geographic region, and teaching status are assigned according to information from the American Hospital Association Annual Survey of Hospitals. The proportion of uninsured and Medicaid patients admitted to each hospital was found to have a nonmonotonic relationship with the outcomes being assessed and was thus treated as a 6‐level categorical variable with the following levels: 0% to 10%, 11% to 20%, 21% to 30%, 31% to 40%, 41% to 50%, and 51% to 100%.

Statistical Analysis

Summary statistics were constructed at the patient level and differences in proportions were evaluated with the chi‐square test. We employed direct standardization, using the age and sex distribution of the entire cohort, to compute age‐standardized and sex‐standardized estimates for each insurance group and compared them using the chi‐square test for in‐hospital mortality and t test for log transformed LOS and cost per hospitalization. For each condition, we developed multivariable logistic regression models for in‐hospital mortality and multivariable linear regression models for log transformed LOS and cost. The patient was the unit of analysis in all models.

In order to elucidate the contribution of disease severity and comorbidities and the proportion of uninsured and Medicaid patients admitted to each hospital, we fitted 3 sequential models for each outcome measure: Model 1 adjusted for patient sociodemographic characteristics and hospital characteristics with the exception of the covariate for the proportion of uninsured and Medicaid patients, Model 2 adjusted for all covariates in the preceding model as well as patients' comorbidities and severity of principal diagnosis, and Model 3 adjusted for all previously mentioned covariates as well as the proportion of uninsured and Medicaid patients admitted to each hospital. We excluded patients who died during hospitalization from the models for LOS and cost. We exponentiated the effect estimates from the log transformed linear regression models so that the adjusted ratio represents the percentage change in the mean LOS or mean cost.

To determine whether the association between patients' insurance status and in‐hospital mortality was modified by seeking care in hospitals treating a smaller or larger proportion of uninsured and Medicaid patients, we entered an interaction term for insurance status and proportion of uninsured and Medicaid patients in the final models (Model 3) for our primary outcome of in‐hospital mortality. However, since no significant interaction was found for any of the 3 conditions, this term was removed from the models and results from the interaction models are not described. In order to assess model specification for the linear regression models, we evaluated the normality of model residuals and found that these were approximately normally distributed. Lastly, we attempted to test the robustness of our analyses by creating fixed effects models that controlled for hospital site but were unable to do so due to the computational limitations of available software packages that could not render fixed effects models with more than 1000 hospital sites.

For all variables except race/ethnicity, data were missing for less than 3% of patients, so we excluded these individuals from adjusted analyses. However, race/ethnicity data were missing for 29% of the sample and were analyzed in 3 different ways, namely, with the missing data treated as a separate covariate level, with the missing data removed from the analysis, and with the missing data assigned to the majority covariate level (white race). The results of our analysis were unchanged no matter how the missing values were assigned. As a result, missing values for race/ethnicity were treated as a separate covariate level in the final analysis.15 Sociodemographic characteristics of patients with missing race/ethnicity information were similar to those with complete data.

We used SUDAAN (Release 9.0.1, Research Triangle Institute, NC) to account for NIS's sampling scheme and generalized estimating equations to adjust for the clustering of patients within hospitals and hospitals within sampling strata.29 In order to account for NIS's stratified probability sampling scheme, SUDAAN uses Taylor series linearization for robust variance estimation of descriptive statistics and regression parameters.30, 31 We present 2‐tailed P values or 95% confidence intervals (CIs) for all statistical comparisons.

Patient and Hospital Characteristics

The final cohort comprised of 154,381 patients discharged from 1018 hospitals in 37 states during calendar year 2005 (Table 1). This cohort was representative of 755,346 working‐age Americans, representing approximately 225,947 cases of AMI (29.9%), 151,812 cases of stroke (20.1%), and 377,588 cases of pneumonia (50.0%). Of these patients, 47.5% were privately insured, 12.0% were uninsured, 17.0% received Medicaid, and 23.5% were assigned to Medicare. Compared with the privately insured, uninsured and Medicaid patients were generally younger, less likely to be white, more likely to have lower income, and more likely to be admitted through the emergency department. Of the 1018 hospitals included in our study, close to half (44.3%) were small, with bedsize ranging from 24 to 249. A large number of hospitals were located in the South (39.9%), and 14.9% were designated teaching hospitals.

Compared with privately insured patients, a larger proportion of uninsured and Medicaid patients had higher predicted mortality levels (Table 2). Medicaid patients had a disproportionately higher predicted LOS, whereas predicted resource demand was higher among privately insured patients. Hypertension (48%), chronic pulmonary disease (29.5%), and uncomplicated diabetes (21.5%) were the 3 most common comorbidities in the study cohort, with a generally higher prevalence of comorbidities among Medicaid patients.

In‐Hospital Mortality

Compared with the privately insured, age‐standardized and sex‐standardized in‐hospital mortality for AMI and stroke was significantly higher for uninsured and Medicaid patients (Table 3). Among pneumonia patients, Medicaid recipients had significantly higher in‐hospital mortality compared with privately insured and uninsured patients.

After multivariable adjustment for additional patient and hospital characteristics, uninsured AMI and stroke patients continued to have significantly higher in‐hospital mortality compared with the privately insured (Table 4). Among pneumonia patients, Medicaid recipients persisted in having significantly higher in‐hospital mortality than the privately insured.

LOS

Among AMI and stroke patients, age‐standardized and sex‐standardized mean LOS was significantly longer for the uninsured and Medicaid recipients compared with the privately insured (Table 3). Among pneumonia patients, the uninsured had a slightly shorter mean LOS compared with the privately insured whereas Medicaid recipients averaged the longest LOS.

These insurance‐related disparities in LOS among pneumonia patients persisted after multivariable adjustment (Table 4). Among AMI patients, only Medicaid recipients persisted in having a significantly longer LOS than the privately insured. Among stroke patients, both the uninsured and Medicaid recipients averaged a longer LOS compared with the privately insured.

Cost per Episode

For all 3 conditions, the uninsured had significantly lower age‐standardized and sex‐standardized costs compared with the privately insured (Table 3). However, Medicaid patients had higher costs than the privately insured for all three conditions, significantly so among patients with stroke and pneumonia.

These insurance‐related disparities in costs persisted in multivariable analyses (Table 4). The uninsured continued to have lower costs compared with the privately insured, significantly so for patients with AMI and pneumonia. Among stroke and pneumonia patients, Medicaid recipients continued to accrue higher costs than the uninsured or privately insured.

Limitations

Our study has several limitations. Since the NIS does not assign unique patient identifiers that would permit tracking of readmissions, we excluded patients transferred to another acute‐care hospital from our study to avoid counting the same patient twice. However, only 10% of hospitalized patients underwent transfer for cardiac procedures in the National Registry of Myocardial Infarction, with privately insured patients more likely to be transferred than other insurance groups.17 Since these patients are also more likely to have better survival, their exclusion likely biased our study toward the null. The same is probable for stroke patients as well.

Some uninsured patients begin Medicaid coverage during hospitalization and should ideally be counted as uninsured but were included under Medicaid in our analysis. They are also likely to be state‐ and plan‐specific variations in Medicaid and private payer coverage that we could not incorporate into our analysis. In addition, we were unable to include deaths that may have occurred shortly after discharge, even though these may have been related to the quality of hospital care. Furthermore, although the 3 conditions we studied are common and responsible for a large number of hospital deaths, they make up about 8% of total annual hospital discharges,23 and caution should be exercised in generalizing our findings to the full spectrum of hospitalizations. Lastly, it is possible that unmeasured confounding could be responsible for the observed associations. Uninsured and Medicaid patients are likely to have more severe disease, which may not be adequately captured by the administrative data available in the NIS. If so, this would explain the mortality association rather than insurance status.36, 37

Design and Subjects

We examined data from the 2005 Nationwide Inpatient Sample (NIS), a nationally representative database of hospital inpatient stays maintained by the Agency for Healthcare Research and Quality (AHRQ) as part of the Healthcare Cost and Utilization Project (HCUP).20, 21 The NIS is a stratified probability sample of 20% of all US community hospitals, including public hospitals, academic medical centers, and specialty hospitals. Long‐term care hospitals, psychiatric hospitals, and alcoholism/chemical‐dependency treatment facilities are excluded. The 2005 NIS contains data on 7,995,048 discharges from 1054 hospitals located in 37 States and is designed to be representative of all acute care discharges from all US community hospitals.21

We identified discharges with a principal diagnosis of acute myocardial infarction (AMI), stroke, and pneumonia using International Classification of Diseases, 9th Revision, Clinical Modification (ICD‐9‐CM) codes specified in the AHRQ definitions of Inpatient Quality Indicators (Supporting Information Appendix).22 These 3 conditions are among the leading causes of noncancer inpatient deaths in patients under 65 years old,23 and evidence suggests that high mortality may be associated with deficiencies in the quality of inpatient care.24

We confined our analysis to patients 18 to 64 years of age, since this population is most at risk of being uninsured or underinsured.25 We excluded pregnant women because they account for an unusually high proportion of uninsured discharges and were relatively few in our cohort.26 In addition, we excluded patients transferred to another acute care hospital and patients missing payer source and discharge disposition. Our study protocol was approved by the Partners Human Research Committee.

Study Variables

We categorized insurance status as privately insured, uninsured, Medicaid, or Medicare. We defined privately insured patients as those having either Blue Cross or another commercial carrier listed as the primary payer and uninsured patients as those having either no charge or self‐pay listed as the primary payer.27 Other governmental payer categories were noted to share several characteristics with Medicare patients and comprised only a small proportion of the sample, and were thus included with Medicare. In order to account for NIS's sampling scheme and accurately apply sample weights in our analysis, we used Medicare as a separate category. However, since Medicare patients age 18 to 64 years represent a fundamentally different population that is primarily disabled or very ill, only results of privately insured, uninsured, and Medicaid patients are reported.

We selected in‐hospital mortality as the outcome measure and LOS and cost per hospitalization as measures of resource utilization. The NIS includes a binary indicator variable for in‐hospital mortality and specifies inpatient LOS in integers, with same‐day stays coded as 0. NIS's cost estimates are based on hospital cost reports submitted to the Centers for Medicare and Medicaid Services. To test the validity of our cost analyses, we performed parallel analyses using hospital charges as a measure of utilization (charges include hospital overhead, charity care, and bad debt). The resulting adjusted ratios differed little from cost ratios and we opted to report only the details of our cost analyses.

In order to assess the independent association between insurance status and the outcome measures listed above, we selected covariates for inclusion in multivariable models based on the existing literature. Patient covariates included: age group (18‐34 years, 35‐49 years, 50‐64 years), sex (male/female), race/ethnicity (non‐Hispanic white, non‐Hispanic black, Hispanic, other, missing), median income by zip code of residence (less than $37,000, $37,000‐$45,999, $46,000‐$60,999, $61,000 or more), admission through the emergency department (yes/no), admission on a weekend (yes/no), measures of disease severity, and comorbidity indicators. Measures of disease severity specific to each outcome are assigned in the NIS using criteria developed by Medstat (Medstat Disease Staging Software Version 5.2, Thomson Medstat Inc., Ann Arbor, MI). Severity is categorized into 5 levels, with a higher level indicating greater risk. We recorded comorbidities for each patient in our sample using HCUP Comorbidity Software, Version 3.2 (www.hcup‐us.ahrq.gov/tools_software.jsp) based on comorbidity measures reported by Elixhauser et al.28

Hospital covariates included: bed size (small, medium, large), ownership/control (private, government, private or government), geographic region (northeast, midwest, south, west), teaching status (teaching, non‐teaching), and the proportion of uninsured and Medicaid patients (combined) admitted to each hospital for AMI, stroke, or pneumonia. The actual number of hospital beds in each bed size category varies according to a hospital's geographic region and teaching status.27 Ownership/control, geographic region, and teaching status are assigned according to information from the American Hospital Association Annual Survey of Hospitals. The proportion of uninsured and Medicaid patients admitted to each hospital was found to have a nonmonotonic relationship with the outcomes being assessed and was thus treated as a 6‐level categorical variable with the following levels: 0% to 10%, 11% to 20%, 21% to 30%, 31% to 40%, 41% to 50%, and 51% to 100%.

Statistical Analysis

Summary statistics were constructed at the patient level and differences in proportions were evaluated with the chi‐square test. We employed direct standardization, using the age and sex distribution of the entire cohort, to compute age‐standardized and sex‐standardized estimates for each insurance group and compared them using the chi‐square test for in‐hospital mortality and t test for log transformed LOS and cost per hospitalization. For each condition, we developed multivariable logistic regression models for in‐hospital mortality and multivariable linear regression models for log transformed LOS and cost. The patient was the unit of analysis in all models.

In order to elucidate the contribution of disease severity and comorbidities and the proportion of uninsured and Medicaid patients admitted to each hospital, we fitted 3 sequential models for each outcome measure: Model 1 adjusted for patient sociodemographic characteristics and hospital characteristics with the exception of the covariate for the proportion of uninsured and Medicaid patients, Model 2 adjusted for all covariates in the preceding model as well as patients' comorbidities and severity of principal diagnosis, and Model 3 adjusted for all previously mentioned covariates as well as the proportion of uninsured and Medicaid patients admitted to each hospital. We excluded patients who died during hospitalization from the models for LOS and cost. We exponentiated the effect estimates from the log transformed linear regression models so that the adjusted ratio represents the percentage change in the mean LOS or mean cost.

To determine whether the association between patients' insurance status and in‐hospital mortality was modified by seeking care in hospitals treating a smaller or larger proportion of uninsured and Medicaid patients, we entered an interaction term for insurance status and proportion of uninsured and Medicaid patients in the final models (Model 3) for our primary outcome of in‐hospital mortality. However, since no significant interaction was found for any of the 3 conditions, this term was removed from the models and results from the interaction models are not described. In order to assess model specification for the linear regression models, we evaluated the normality of model residuals and found that these were approximately normally distributed. Lastly, we attempted to test the robustness of our analyses by creating fixed effects models that controlled for hospital site but were unable to do so due to the computational limitations of available software packages that could not render fixed effects models with more than 1000 hospital sites.

For all variables except race/ethnicity, data were missing for less than 3% of patients, so we excluded these individuals from adjusted analyses. However, race/ethnicity data were missing for 29% of the sample and were analyzed in 3 different ways, namely, with the missing data treated as a separate covariate level, with the missing data removed from the analysis, and with the missing data assigned to the majority covariate level (white race). The results of our analysis were unchanged no matter how the missing values were assigned. As a result, missing values for race/ethnicity were treated as a separate covariate level in the final analysis.15 Sociodemographic characteristics of patients with missing race/ethnicity information were similar to those with complete data.

We used SUDAAN (Release 9.0.1, Research Triangle Institute, NC) to account for NIS's sampling scheme and generalized estimating equations to adjust for the clustering of patients within hospitals and hospitals within sampling strata.29 In order to account for NIS's stratified probability sampling scheme, SUDAAN uses Taylor series linearization for robust variance estimation of descriptive statistics and regression parameters.30, 31 We present 2‐tailed P values or 95% confidence intervals (CIs) for all statistical comparisons.

Patient and Hospital Characteristics

The final cohort comprised of 154,381 patients discharged from 1018 hospitals in 37 states during calendar year 2005 (Table 1). This cohort was representative of 755,346 working‐age Americans, representing approximately 225,947 cases of AMI (29.9%), 151,812 cases of stroke (20.1%), and 377,588 cases of pneumonia (50.0%). Of these patients, 47.5% were privately insured, 12.0% were uninsured, 17.0% received Medicaid, and 23.5% were assigned to Medicare. Compared with the privately insured, uninsured and Medicaid patients were generally younger, less likely to be white, more likely to have lower income, and more likely to be admitted through the emergency department. Of the 1018 hospitals included in our study, close to half (44.3%) were small, with bedsize ranging from 24 to 249. A large number of hospitals were located in the South (39.9%), and 14.9% were designated teaching hospitals.

Compared with privately insured patients, a larger proportion of uninsured and Medicaid patients had higher predicted mortality levels (Table 2). Medicaid patients had a disproportionately higher predicted LOS, whereas predicted resource demand was higher among privately insured patients. Hypertension (48%), chronic pulmonary disease (29.5%), and uncomplicated diabetes (21.5%) were the 3 most common comorbidities in the study cohort, with a generally higher prevalence of comorbidities among Medicaid patients.

In‐Hospital Mortality

Compared with the privately insured, age‐standardized and sex‐standardized in‐hospital mortality for AMI and stroke was significantly higher for uninsured and Medicaid patients (Table 3). Among pneumonia patients, Medicaid recipients had significantly higher in‐hospital mortality compared with privately insured and uninsured patients.

After multivariable adjustment for additional patient and hospital characteristics, uninsured AMI and stroke patients continued to have significantly higher in‐hospital mortality compared with the privately insured (Table 4). Among pneumonia patients, Medicaid recipients persisted in having significantly higher in‐hospital mortality than the privately insured.

LOS

Among AMI and stroke patients, age‐standardized and sex‐standardized mean LOS was significantly longer for the uninsured and Medicaid recipients compared with the privately insured (Table 3). Among pneumonia patients, the uninsured had a slightly shorter mean LOS compared with the privately insured whereas Medicaid recipients averaged the longest LOS.

These insurance‐related disparities in LOS among pneumonia patients persisted after multivariable adjustment (Table 4). Among AMI patients, only Medicaid recipients persisted in having a significantly longer LOS than the privately insured. Among stroke patients, both the uninsured and Medicaid recipients averaged a longer LOS compared with the privately insured.

Cost per Episode

For all 3 conditions, the uninsured had significantly lower age‐standardized and sex‐standardized costs compared with the privately insured (Table 3). However, Medicaid patients had higher costs than the privately insured for all three conditions, significantly so among patients with stroke and pneumonia.

These insurance‐related disparities in costs persisted in multivariable analyses (Table 4). The uninsured continued to have lower costs compared with the privately insured, significantly so for patients with AMI and pneumonia. Among stroke and pneumonia patients, Medicaid recipients continued to accrue higher costs than the uninsured or privately insured.

Limitations

Our study has several limitations. Since the NIS does not assign unique patient identifiers that would permit tracking of readmissions, we excluded patients transferred to another acute‐care hospital from our study to avoid counting the same patient twice. However, only 10% of hospitalized patients underwent transfer for cardiac procedures in the National Registry of Myocardial Infarction, with privately insured patients more likely to be transferred than other insurance groups.17 Since these patients are also more likely to have better survival, their exclusion likely biased our study toward the null. The same is probable for stroke patients as well.

Some uninsured patients begin Medicaid coverage during hospitalization and should ideally be counted as uninsured but were included under Medicaid in our analysis. They are also likely to be state‐ and plan‐specific variations in Medicaid and private payer coverage that we could not incorporate into our analysis. In addition, we were unable to include deaths that may have occurred shortly after discharge, even though these may have been related to the quality of hospital care. Furthermore, although the 3 conditions we studied are common and responsible for a large number of hospital deaths, they make up about 8% of total annual hospital discharges,23 and caution should be exercised in generalizing our findings to the full spectrum of hospitalizations. Lastly, it is possible that unmeasured confounding could be responsible for the observed associations. Uninsured and Medicaid patients are likely to have more severe disease, which may not be adequately captured by the administrative data available in the NIS. If so, this would explain the mortality association rather than insurance status.36, 37