Inpatient Childhood Asthma Treatment: Title and subTitle BreakRelationship of Hospital Characteristics to Length of Stay and Cost: Analyses of New York State Discharge Data, 1995 FREE

ASTHMA IS the leading cause of chronic illness in children and young adults and one of the most frequent causes of hospitalization.^{1 - 4} Furthermore, pediatric asthma hospitalization rates have increased in the past 2 decades.^{2 - 5} Asthma alone accounts for 1% of all health care expenditures in the nation,⁶ with a substantial portion of this cost derived from inpatient care.⁷ Therefore, it is increasingly important to optimize inpatient treatment, and one approach to doing so is to examine institutional variability in efficiency in inpatient treatment of asthma in pediatric patients.

Available evidence is limited regarding the relationship between hospital type and efficiency of care for children. Teaching hospitals perform most of the charity care in the United States⁸ and academic physicians care for a higher percentage of uninsured patients.⁹ Pediatric patients with asthma who are uninsured, Medicaid-insured, or live in neighborhoods with lower socioeconomic status (SES) are more severely ill, have longer length of stay (LOS), and higher hospital costs.¹⁰ Studies of LOS and hospital charges in teaching hospitals have yielded contradictory findings regarding efficiency after adjusting for case mix or severity of illness.^{11 - 13} Nevertheless, it has been suggested that health plans have been reluctant to contract with teaching hospitals for all but the most complex care because they perceive teaching hospitals to be too costly and inefficient.¹⁴

Similarly, public hospitals have a unique and important role in the health care market.¹⁵ In particular, New York State has a large number of public hospitals providing "safety net care" for disadvantaged populations.¹⁵ Few studies have compared their performance with that of private hospitals. Concerning private hospitals, many assume that the growth of for-profit hospitals is attributable to lower costs and greater efficiency. However, administrative costs are proportionately higher in private for-profit hospitals.¹⁶

To date, there is a single report, by Meurer et al,¹² comparing hospital efficiency in treating pediatric asthma in different types of hospitals. Although this important study found that mean charges for childhood asthma varied significantly by hospital ownership and teaching status, the authors analyzed data from the Healthcare Cost and Utilization Project, with oversampling of small teaching hospitals and large nonteaching hospitals. Individual hospital charge-to-cost conversion ratios and patient ethnicity data were not available for adjustment. No specific statistical procedure was used to control for the clustering of relevant patient characteristics within hospitals.

Analyses reported here examine the association of selected hospital characteristics (teaching status and hospital ownership) with cost and LOS among pediatric patients admitted for asthma. We linked a unique population-based discharge data set from the state of New York to census and American Hospital Association data and used statistical modeling to adjust for the effect of clustered patient characteristics in each hospital. Using these methods, we tested the hypothesis that hospital costs and LOS for children with asthma are associated with hospital type.

We analyzed the 1995 Statewide Planning and Research Cooperative System (SPARCS) database, describing all inpatient discharges from New York State acute care hospitals. Hospital information from the discharge data abstract and the uniform billing form completed by billing departments are merged to create SPARCS. Duplicate records and missing data are identified and corrected.¹⁷ The database includes demographic, diagnostic, utilization, and financial information. Primary diagnosis, up to 14 secondary diagnoses, procedures, patient insurance status, and the Permanent Facility Identifier for hospitals are included. American Hospital Association data and the appropriate census data¹⁸ were linked to 1995 SPARCS by the Permanent Facility Identifier and patient ZIP code, respectively. This provided information on hospital type and socioeconomic indicators in the ZIP code of residence for each discharged patient.

The Children's National Medical Center institutional review board granted an exemption from review based on the anonymous nature of the data and their public availability.

Patients were included in this study if they were aged 1 to 17 years with a primary diagnosis of asthma (International Classification of Diseases, Ninth Revision¹⁹ codes 493.00-493.91). To construct models that best represented most asthma discharges, we excluded outliers with total cost or LOS greater than the 98th percentile and those with cost less than the second percentile.

The outcome measures were total cost and LOS. The total 1995 dollar cost for each discharge was calculated from the total charges by application of the hospital-specific overall charge-to-cost ratio. This information is provided by each hospital in the state of New York and available as part of selected versions of the SPARCS data set. Length of stay was recorded in SPARCS in full-day units. Hospital characteristics, such as ownership (public/private) were abstracted from the American Hospital Association database. Teaching hospitals were defined by (1) membership in the Council of Teaching Hospitals of the Association of American Medical Colleges or (2) residency training approved by the Accreditation Council for Graduate Medical Education.²⁰

There are obvious difficulties with developing severity measures in administrative databases.²¹ For this reason, a standard method with accepted definitions but well-recognized limitations was used: the All Patients Refined Diagnosis Related Group (APR-DRG), which was applied by using standard software developed by 3M Information Systems (Salt Lake City, Utah) and the National Association of Children's Hospital and Related Institutions (Alexandria, Va). The severity of APR-DRG is based on secondary diagnoses and categorizes patients from low to high severity by placing them in 1 of 4 categories.^{22 - 23}

Age in years, sex, and race (white, black, other) were recorded in SPARCS. Insurance type was classified into 3 groups: Medicaid, commercial, and other.

To estimate individual socioeconomic status (not available in SPARCS), the median household income by ZIP code was retrieved from 1990 census data.¹⁸ Zip codes were then categorized into 4 groups based on the median household income according to the Healthcare Cost and Utilization Project database: (1) less than $25 000; (2) $25 001 to $30 000; (3) $30 001 to $35 000; (4) greater than $35 000. The validation and usefulness of using census data in this way has been documented in several studies.²⁴ Krieger²⁴ suggested that such census data should be used only for analyses occurring within 5 years of the census because population growth and migration alter a neighborhood's composition. The data we used met this criterion.

We first examined whether severity differed by hospital type. For each hospital characteristic, separate tables were produced examining severity by hospital type (teaching vs nonteaching and private vs public). The statistical significance of the distribution of severity within each hospital type and each of the patient characteristics was assessed by a χ² test. Mean values of the dependent variables (cost and LOS) were also calculated by hospital type and by patient characteristics. These variables were not normally distributed. For this reason as well as to reduce the chance of a type I error, we chose to use bootstrap confidence intervals for the within-variable mean comparisons. These intervals use the bias-corrected and accelerated (BC_a) method, which is preferred when normality is questionable.²⁵

The overall effect of the hospital characteristics on LOS and cost, while adjusting for patient characteristics, was evaluated by a series of regressions. When examining data collected on more than 1 level (eg, hospital and patient) there is the chance that the SEs of the parameter estimates are biased. Hierarchical models (also called random coefficients, mixed models, and multilevel models) are often used.²⁶ These models adjust the SEs and allow for correct test statistics to be calculated. We implemented these models by using SAS PROC MIXED (general linear mixed models) (SAS Institute, Cary, NC) for the cost outcome and the SAS GLIMMIX macro (generalized linear mixed models) for the Poisson regressions on the LOS outcome. These models are covariance-adjusted for the hospital-level characteristics by using the hospital identifier as the clustering variable. Covariance adjustment involves using models that account for within- and between-hospital variation. We compared models that adjust for this variance with standard regression models to assess the need for this type of analysis. All statistical analyses were performed using SAS version 8.1.

The 1995 SPARCS data set contains 19 688 discharges of patients between the ages of 1 and 17 years with a primary diagnosis of asthma, representing discharges from 206 hospitals in New York State. Nearly half of these institutions were teaching hospitals and only 12% were public hospitals (Table 1). As presented in Table 2, most patients were cared for in teaching hospitals (89%) and privately owned hospitals (69%).

Table 3 presents selected patient characteristics in different types of hospitals. There was little difference in APR-DRG severity by hospital type. Public and teaching hospitals had more patients insured by Medicaid and fewer commercially insured. Age distributions were similar in different types of hospitals but there were uniformly more boys (60%-62%) than girls. Most patients in public hospitals were of races listed as "other than white or black," but in private hospitals race was distributed more uniformly. Teaching hospitals had more minority patients than nonteaching hospitals, in which most were white (76%). Patients from poorer neighborhoods tended to be discharged from public and teaching hospitals.

Unadjusted mean costs differed by hospital characteristics (Table 4). Mean cost was significantly but trivially higher in private hospitals ($1868 vs $1771) and in nonteaching hospitals ($1876 vs $1528). Mean LOS (2.07 days) did not differ significantly by hospital characteristics.

Regression models assessed mean LOS and cost by each hospital characteristic after adjusting for the other hospital characteristic and again after adjusting for both the other hospital characteristic and patient characteristics. The hospital level variance explained 17% of the total variability in cost but only 4% of the overall variability in LOS. Using Akaike's²⁷ information criteria, an informal method of examining the goodness of fit between adjusted and unadjusted models, there is evidence that the adjusted models do improve fit. However, the fact that only 4% of the overall variance in LOS is explained by the within-hospital correlation suggests that the multilevel model does not substantially improve explained variability in this outcome.

As presented in Table 5, after adjusting for the patient-level data and the effect of the hierarchical nature of the data, there are not meaningful differences in LOS or cost among different types of hospitals. On the other hand, because of the large population, some of these differences achieve statistical significance (eg, teaching hospitals cost $188 more per day).

Although analyses excluded LOS and cost outliers, other patient characteristics could have affected the results. Therefore, additional analyses excluded patients who were transferred or who had severe secondary diagnoses (such as congenital heart disease or cystic fibrosis) identified by a consensus panel of acute care pediatricians. There were only 584 patients (2.9%) with severe comorbidities and 548 patients (2.8%) who were transferred. The main results (Table 5) did not differ in these supplementary analyses and are therefore not reported further.

These findings provide little empiric evidence for a relationship between resource utilization and hospital type. Unlike previous studies, we applied statistical methods, such as hierarchical modeling and bootstrap confidence intervals, to account for the clustering effect of hospitals and nonnormal distribution of the outcomes.

Almost one third of pediatric patients with asthma were treated in public hospitals and, not unexpectedly, were disproportionately drawn from socioeconomically disadvantaged areas, were nonwhite, and were insured through Medicaid. Given these facts, it becomes especially important to consider the efficiency of public hospitals. This large, comprehensive administrative database revealed that public hospitals performed as well as private hospitals in mean length and cost of pediatric asthma admissions, suggesting that these safety net hospitals provided care for large numbers of disadvantaged children with asthma in an equivalently efficient manner.

A recent study also failed to find a significant difference in average LOS by hospital ownership.¹² Average charges were higher in private for-profit hospitals compared with public hospitals, but no difference was found between private nonprofit hospitals and public hospitals. In New York State in 1995, the small number of childhood asthma discharges from private for-profit hospitals (43 discharges [0.02%]) made analysis of such hospitals impossible. Almost all private hospitals in our study were nonprofits and thus, our finding is consistent with this previous report.

Although there were approximately equal numbers of teaching hospitals and nonteaching hospitals in New York State, teaching hospitals accounted for nearly 90% of the pediatric discharges and these children were more frequently Medicaid-insured. After adjustment for other covariates and the hierarchical nature of the data, neither LOS nor cost was meaningfully different in hospitals training pediatricians.

It is interesting that cost was lower in teaching hospitals before adjustment for other hospital characteristics but higher when adjusted. This is probably because more teaching hospitals were public and cost in public hospitals was lower than in private hospitals. When adjusted for ownership, the cost became higher in teaching hospitals.

The relationship between hospital type and efficiency of treatment has been examined in adult populations, but far less is known regarding this important issue in children. Many analyses of adult conditions have dealt with technologically complex or procedure-driven treatments, such as coronary angioplasty.^{28 - 29} We focused on asthma, the most common condition requiring hospitalization in children. Our results do not reflect efficiencies in specific procedures or technologies but are potentially more relevant to multiple and diverse diagnoses.

New York State data were analyzed for a variety of reasons, including the diversity of the population and of the hospitals. It is unknown whether these results can be generalized to other states and regions. Further, we recognize that LOS in New York State is known to be generally longer than many other areas.^{30 - 31} Our choice of 1995 data reflected the need to validly link to the 1990 census data and the desire to analyze information obtained prior to managed care penetration. Local markets in a large and heterogeneous state may change in relatively unpredictable and variable fashion early in managed care penetration. We believe a more thorough analysis of this issue in its own right is required rather than confounding analyses of hospital characteristics with the equally complex issue of managed care effects. Nonetheless, taken together, these limitations must be considered when interpreting the analyses.

Our analyses are incomplete in several ways. Most notably, in common with all administrative data, they do not include detailed patient or hospital information. For example, there is no clinical information that might provide a more detailed severity adjustment, and the databases do not include factors such as staff ratios and labor costs in the participating hospitals. For these reasons, we chose a straightforward and descriptive interpretation rather than speculating well beyond the scope of the data.

Another limitation of this administrative database is that no unique patient identifier was available and therefore repeated admissions cannot be traced. Using the modest data available, a proxy identifier was constructed using race, sex, age, insurance type, and residence ZIP code for each discharge in our study. Using this measure, we determined that 18% of discharges could be accounted for by readmission. However, due to the obvious limitation of the estimation, we did not use it in our hypothesis testing or account for the "admission clustering," by which 1 admission raises the probability of another. Furthermore, in common with other analyses on administrative data, the inability to sort out readmissions creates theoretical difficulties concerning the statistical assumption of independent observations.

In the face of these limitations, we nonetheless documented that the efficiency in hospitals that provide medical education and safety net care is equivalent to that found in nonteaching and privately owned hospitals. Full discussion of these results requires consideration of much broader but pressing issues. In particular, how does one value the training of future pediatricians or caring for disadvantaged children? Is it even reasonable to use efficiency to determine support for both training and safety net care in our currently complex and evolving health care market? It is our hope that such important and on-going policy discussions can be informed by the analyses provided in this article.

Using a large administrative database for the state of New York in 1995, we found that public hospitals delivered asthma hospital care to disadvantaged patients. Using appropriate statistical modeling, the cost and LOS for care in public hospitals was not different from that in private hospitals. There were statistically significant but trivial differences in the care provided by teaching hospitals, with mean cost being slightly higher and mean LOS, lower. These results have potentially important implications for the organization of health care delivery in the United States.

Accepted for publication September 23, 2001.

This study was presented in part at the annual meeting of the Ambulatory Pediatric Association, San Francisco, Calif, May 1, 1999.

Corresponding author and reprints: Jill G. Joseph, MD, PhD, Center for Health Services and Clinical Research, Children's National Medical Center, 111 Michigan Ave NW, Washington, DC 20010 (e-mail: jjoseph@cnmc.org).

Asthma is the most common admitting diagnosis in children and there is increasing pressure to optimize treatment efficiency. It is possible that hospital characteristics influence such efficiency. Using advanced statistical models, we found statistically significant but trivial differences in the care provided by teaching hospitals, with mean cost being slightly higher. On the other hand, public hospitals provided care that did not differ in either cost or LOS from that provided in private hospitals. Thus, hospitals providing both training for future pediatricians and safety net care for disadvantaged children treat pediatric asthma without increasing LOS or cost.