Objectives. To examine geographic variation in rates of infant hospitalization for diagnoses classified by type of hospitalization decision in Monroe County (Rochester), New York.
Methods. Study design was cross-sectional and ecologic.International Classification of Diseases (ICD) codes were used to categorize all 7883 hospitalizations for infants (age, <24 months) beyond the newborn period between 1985 and 1991. Postal zip codes defined socioeconomic areas as inner-city, other urban, and suburban for the population at risk. In 1990, inner-city infants included 62% black and 65% Medicaid-covered infants, whereas suburban infants included 3% black and 6% covered by Medicaid. Hospitalization rates were compared among the three socioeconomic areas.
Results. Overall hospitalization rate was 50.3 per 1000 child years. Admissions classified as discretionary accounted for 59% of these, followed by those classified as mandatory, 18%; sometime (congenital heart disease, cleft palate), 15%; discretionary surgery (inguinal hernia, tonsillectomy/adenoidectomy), 6%; and unlikely to need admission, 2%. A stepwise, socioeconomic gradient in hospitalization was found, with rates of 38.1, 51.3, and 82.9 per 1000 child-years, respectively, for suburban, other urban, and inner-city areas. Rates for discretionary, unlikely, and mandatory admissions followed this gradient. Using the odds for hospitalization of suburban infants as the base odds, the odds ratio for discretionary hospitalization for inner-city infants was 2.88 (95% confidence interval [CI], 2.69 to 3.08) and that for mandatory hospitalization was 2.20 (95% CI, 1.94 to 2.49). In multiple regression analysis, low education level of mothers explained 81% of the variance in discretionary hospitalization rate. Although the per capita rate of hospital care of inner-city infants was more than twofold greater than that for suburban infants, potential for reducing this difference is suggested by the fact that discretionary admissions accounted for 78.9% of this difference, whereas mandatory admissions accounted for 17.7% of the difference.
Conclusion. The hospitalization rate for inner-city infants is much greater than that for suburban infants. A substantial portion of the difference, namely that attributable to mandatory admissions, reflected higher rates of serious illness. Differences attributable to discretionary admissions may reflect higher rates of serious illness to some extent, but also appear to reflect less effective health services to a substantial degree.
Hospitalization accounts for almost 50% of child health expenses,1 estimated at $49.8 billion per year for 1987.2 Admission rate, not length of stay, is the primary determinant of per capita hospital use.3 Studies on geographic variation in rates of pediatric hospitalization suggest that a large proportion of admissions are discretionary, and that much opportunity for reducing hospitalization exists.3-8Admitting physicians in an urban teaching hospital judged 28% of pediatric medical admissions to be potentially avoidable.9Some conditions commonly leading to hospitalization in children, such as gastroenteritis and asthma, have been widely recognized as sensitive to the quality of primary health care.10-14
Variation in hospitalization rates for children has been attributed to area-specific physician practice patterns15-22 and differences in hospital bed supply.23 Differences in primary care provided to different socioeconomic groups has received relatively little attention as an explanation for geographic variations in hospitalization rates. Our recent study of infants hospitalized in Monroe County (Rochester), New York, however, documented more than threefold greater rates in the inner city than the suburbs for lower respiratory tract illness and acute febrile illness.7 The vast majority (95%) of these hospitalizations occurred at two hospitals whose emergency departments are both staffed by a single group of pediatric attending and resident physicians. Because these two hospitals function as a single supply of hospital beds, variation cannot be attributed to differences in bed supply.
The present small-area variation analysis also compares hospitalization rates among different socioeconomic areas within Monroe County, and it contrasts further with previous studies in scope and type of conditions studied. Whereas previous variation studies have generally examined distributions by diagnosis-related groups (DRGs) or have focused on a few specific clinical problems, this study used a classification of hospitalizations that encompassed virtually all hospitalizations among infants. All International Classification of Diseases24 (ICD) diagnoses in hospital discharge files for infants were classified by level of physician discretion in the decision to hospitalize. All hospitalizations were then classified based on the hierarchy among decision classes. ICD rather than DRG codes were used because they allow greater specificity in classifying morbidity, and they are used worldwide.
Specific objectives of this study were: 1) to examine variation among socioeconomic areas in rates of infant hospitalizations classified by this system, and 2) to assess the validity of our classification system in classifying all infant hospitalizations. Infant hospitalizations were chosen because age restriction seemed essential for an initial attempt at reliably classifying a large number of diagnoses, and because a large portion of pediatric admissions occur among infants. Moreover, given the vulnerability of infants and the apparent role of uncertainty in discretionary hospitalization,15,16,25 we expect that discretionary admissions are more prevalent among infants than older children.
Infants were defined by age between 1 and 24 months. The study included all infant hospitalizations for the 7-year period, 1985 through 1991, for residents of Monroe County (Rochester), New York. Using a database including all hospital admissions, maintained by a consortium of Rochester area hospitals throughout the study period, 7883 discharges for infants were recorded. In 1991 and 1992, infants accounted for 29.4% of all hospitalizations for children between the ages of 1 month through 18 years, excluding those related to pregnancy and childbirth. Area of residence was defined by 5-digit postal zip code. For reasons of confidentiality, information to establish residence more precisely was not available. The population at risk for hospitalization was determined from birth certificate data. Denominators, the number of child-years at risk for each year of observation, were estimated as the sum of births for the preceding 2 years. Over the entire study period 156 587 children (child-years) were at risk.
Zip codes and the proportion of births covered by Medicaid were used to divide Monroe County into socioeconomic areas as follows. Inner city was defined as zip codes in which the largest portion of births were covered by Medicaid. Other urban was defined as zip codes with the majority of births to mothers dwelling in the city of Rochester but not in the inner city. Suburbs included remaining zip codes in the county.
Denominators for socioeconomic areas were based on the number of births in areas defined by zip codes. Place of residence by zip code was available in birth data for 2 of the 7 years of observation and by census tract for the entire observation period. Zip code areas and census tracts do not correspond perfectly, but good estimates for the distribution of the births by zip code for the entire observation period were derived from data on the 23 738 births in 1989 and 1990, termed the area definition file, that included information on both zip code area and census tract. The first two rows of Table1 indicate the number of births in each socioeconomic area. Ratios of births by zip code to births by census tract in the area definition file (third row of Table 1) were used to calculate weight factors for converting the number of births in an area as defined by census tracts to the number of births in an area as defined by zip codes. Child-years at risk for areas based on census tracts in the complete birth file (that included 7 years of birth data) were then multiplied by these weight factors to yield estimates for child-years at risk for areas defined by zip codes.
The area definition file was also used to provide information from birth certificates that could be used to characterize the socioeconomic areas defined by zip codes. In addition to births covered by Medicaid, this information included prepaid health insurance; lack of health insurance; time prenatal care was initiated; birth weight; and maternal age, race, ethnicity, educational level, and smoking. Data aggregated by zip codes from the 1990 United States Census also were used to characterize socioeconomic areas. Information obtained from this source included housing units lacking a telephone, housing units lacking an automobile, children under 5 years living in poverty, unemployment rate, and per capita income.
Zip Code Areas
Postal zip codes also were used to define geographic areas that were used as a unit of analysis. Because 5-digit zip codes were available only for the 23 738 births in 1989 and 1990, analysis by zip code was limited to hospitalizations in the years (1990 and 1991) for which birth data provided direct denominator estimates. None of the files contained 9-digit zip codes.
Classification of Diagnoses and Hospitalizations
Diagnoses were classified as mandatory, sometime, discretionary, and unlikely. Mandatory diagnoses were defined as those for an acute condition that is life threatening or has the potential to produce long term disability without (or even with) immediate hospitalization (eg, bacterial meningitis). Withholding hospital admission would very likely be deleterious to the child's health, minimal discretion is involved in the admission decision, and inpatient hospital care is almost always provided. Accordingly, hospitalization rates approximate the incidence of these conditions. Sometime diagnoses included those for which at least one hospitalization would almost always be necessary for diagnosis or management (eg, spina bifida). Discretionary diagnoses are those for which the decision to admit involves a substantial amount of physician judgment—physicians may exercise their discretion to a substantial degree. Discretionary conditions have a wide range in severity and are often managed at home. Conditions generally managed without surgery (eg, umbilical hernia) or with outpatient surgery (eg, inguinal hernia) were also considered discretionary. The final class, unlikely, included diagnoses that are almost always managed without hospitalization and were therefore considered unlikely to need admission. Most of the time, the question of hospitalization never arises for diagnoses in this class (eg, upper respiratory tract infection, acute otitis media)—physician judgment is not involved. When admission does occur for such diagnoses, however, it seems likely that a substantial degree of discretion is exercised.
A hierarchical approach was used to assign each hospitalization to a decision class based on classifications for the first five discharge diagnoses. The four decision classes for diagnoses were ranked, from highest to lowest, as follows: mandatory, sometime, discretionary, and unlikely, and the hospitalization was assigned to the class of the diagnosis with the highest rank. A fifth decision class for hospitalizations was created by using DRG codes to determine whether surgery was performed. If prior classification steps placed the hospitalization in the discretionary or unlikely categories and surgery was performed, the hospitalization was assigned to a fifth decision class termed discretionary surgery.
For the 7883 hospitalizations that were classified, the classification was based on the first diagnosis for 87.5%. An example of a situation in which classification was not based on the first diagnosis were hospitalizations with a first diagnosis of bronchiolitis or asthma (both discretionary) and a subsequent diagnosis of respiratory failure (mandatory).
Development of the Classification
The 1930 different diagnostic codes appearing in the hospital discharge file were grouped in four categories, termed decision classes, that differed in level of physician discretion in the decision for hospitalization. The classification of diagnoses was developed in three stages including: 1) review of literature on small-area variations in hospitalization and other health services, 2) preliminary classification by two investigators (G.S.L., K.M.M.), and 3) final classification based on evaluation of the most common codes by a group of experienced pediatricians.
Existing literature on small-area variations in the use of health service resources provided a conceptual framework and empirical basis for assigning diagnoses to decision classes. Wide variation in per capita rates of health service resource use is generally believed to identify utilization decisions involving high levels of physician discretion; and high rates are believed to reflect overuse of resources.20-22 One view explains small-area variations based on differences in physician response to uncertainty surrounding the effectiveness of medical care. Uncertainty gives rise to divergent practice styles15,16 which, in turn, produce variable rates.17-19
Literature on variations in hospitalization rates supported the validity of assignment of many diagnoses to the discretionary and mandatory classes. Wide variation in hospitalization rates has been observed in multiple studies (listed in Table2) for lower respiratory tract illness (bronchiolitis, asthma, pneumonia), acute febrile illness, gastroenteritis and dehydration, croup, viral meningitis, seizures, urinary tract infections, other gastrointestinal disorders, and skin and soft tissue infections. These seven groups of diagnoses were assigned to the discretionary class. Diagnoses whose assignment was based on prior studies comprised 71.9% of hospitalizations in the discretionary class.
Little variation among broad geographic areas has been observed for several groups of diagnoses fitting the description for mandatory (Table 2), severe fractures (eg, femur), gastrointestinal emergencies (eg, appendicitis), and severe bacterial infections (eg, bacterial meningitis, Haemophilus influenzae type B infections).
Evidence also supported assignment of the majority of diagnoses to the unlikely class. Wide variation in hospitalization rates for such conditions has frequently been observed (Table 2). Diagnoses whose assignment to the unlikely class was empirically based (acute otitis media, upper respiratory tract infection) comprised 88.7% of hospitalizations in the class.
Discretionary surgery, comprising a relatively small proportion (6.2%) of hospitalizations in this community, is difficult to compare with that in DRG-based studies because of the greater specificity achieved with the ICD classification. Two groups of surgical procedures are similar to those previously found to have at least two-fold geographic variations in relative risk, surgery for inguinal hernia, and surgery for tonsil hypertrophy, for adenoid hypertrophy or for chronic middle ear effusion. Inclusion of other groups rests on their content and face validity as surgically-managed conditions that are generally managed without surgery or with outpatient surgery.
Conditions grouped in the sometime class also are difficult to compare with those in DRG-based studies because of the greater specificity achieved with ICD classification. Conditions in the sometime class are almost invariably congenital in origin, and their management generally involves at least one major surgical procedure that requires hospitalization. Although the number of hospital admissions required for each child with one of these conditions may vary with the effectiveness of surgical intervention and with physician discretion, at least one hospital admission is expected for each child. As apparent (Table 2) for 10 groups of disorders accounting for over 90% of admissions in this cluster, content and face validity is high.
In preliminary classification, the 867 specific ICD diagnoses that appeared three or more times among the 7883 admissions were reviewed individually by both these investigators. Discussion over disagreements between independent classification led to refinement of definitions and consensus on final classification for all 867. Based on this experience, the remaining 1063 codes, which had very similar meanings as those already categorized, were classified by one investigator (K.M.M.).
Final classification focused on mandatory and discretionary diagnoses because they were the two most common types, accounting for 76% of all admissions based on preliminary classification, and because these classes seemed to be of greatest importance to child health policy. Altogether, 95 different diagnoses were presented for classification to a panel of 10 experienced, clinically active, hospital-based general pediatricians. Evaluation forms used by panel members included definitions for each class, as above, and offered response options of mandatory, sometime, discretionary, and unlikely. Among the 95 diagnoses were all those classified as mandatory or discretionary in preliminary classification and accounting for 10 or more admissions during the observation period. In addition, six diagnoses with a preliminary classification of sometime and three diagnoses with a preliminary classification of unlikely were presented to encourage the use of all four response options by panel members. The sometime diagnoses evaluated by the panel constituted a 33% random sample of the 18 diagnoses accounting for 10 or more admissions, and the three unlikely diagnoses evaluated included all diagnoses from this class accounting for 10 or more admissions.
Panel members received no training or instructions on classification other than the category descriptions presented above, and there was no interaction among panel members in an attempt to achieve consensus. Diagnoses were ultimately assigned to the class that appeared most frequently among the evaluations by panel members.
Many diagnoses were equivalent to those evaluated by the panel from the perspective of hospitalization decision making, despite ICD codes that were not identical. Such diagnoses were assigned to the same decision classes as the evaluated diagnoses with the same meaning, and they were treated as if they had been evaluated in calculating the proportion of hospitalizations in each class that were attributable to diagnoses evaluated by the panel. For diagnoses not evaluated by the expert panel, the preliminary classification was used as the final classification. The number of diagnoses placed by the panel in each class, and the proportion of all hospitalizations in the class accounted for by these diagnoses, were as follows: discretionary, 116 (87.5%); mandatory 65 (78.0%); sometime, 50 (54.8%); and unlikely, 8 (78.0%). Altogether, 6171 (78.3%) of hospitalizations were attributed to diagnoses classified by the panel, leaving 21.7% for which classification was based the opinion of the authors.
Complete lists of ICD code assignments are available from the authors.
Study design in examining the variation in hospitalization rates among socioeconomic areas was cross-sectional and ecologic. Primary analysis focused on hospitalization rates for three areas (units of analysis) whose characteristics were described, as noted above, on the basis of United States Census data for 1990 and birth certificate data for infants born in 1989 and 1990.
Variations in hospitalization rates among 37 zip code areas were also examined. Correlations and stepwise multiple regressions with stepwise entry of independent variables (listed in Table 1) were used to assess relationships between zip-code specific hospitalization rates for each morbidity cluster and independent variables. Analysis was weighted by the total number of births in each zip code area for 1989 and 1990. Only zip codes for which five or more hospitalizations occurred in 1990 and 1991. Based on this criterion, 2693 of the 2700 hospitalizations in 1990 through 1991 and 37 of Monroe County's 48 zip codes and were included in analysis. Admissions for the 37 zip codes ranged from 8 to 425.
Birth data for 2111 child-years was missing information on zip code. Accounting for only 1.3% of total child-years at risk, these 2111 were included in the denominator only for calculating rates for the entire Monroe County area. The zip code was missing for eight (0.1%) of the hospitalizations. These eight hospitalizations were also excluded from analysis except in calculating rates for the entire county.
Although primary analysis focused on hospitalizations, analysis was also performed in which the event of interest was the hospitalization of a particular child (identified by medical record number) assigned to a specific decision class. The latter events were the basis for calculating patient-class hospitalized rates. In determining these rates, only one hospitalization per decision class (eg, discretionary) was counted for a particular child. Both hospitalization rates and patient-class hospitalized rates are presented as events per 1000 child-years. The calculation of standard errors for rates was based on the binomial distribution.
To further assess the validity of this classification scheme as well as to examine the anticipated socioeconomic disparities in hospitalization rates, several hypotheses reflecting the classification were assessed, as follows: 1) Hospitalization rates for discretionary surgery and sometime classes will vary little among socioeconomic areas. This was anticipated for several reasons. The incidence of sometime conditions, because they are congenital in origin, should vary little between socioeconomic areas. One medical center provides virtually all pediatric subspecialty care in Rochester. Two pediatric surgeons practicing as partners and two pediatric urologists practicing as partners performed the vast majority of pediatric surgery, with the exception of procedures for tonsil or adenoid hypertrophy and chronic middle ear effusion. 2) In moving from suburbs to the other urban and to the inner city, hospitalization rates for the discretionary cluster will increase in stepwise fashion. This hypothesis primarily reflects our belief that city children still receive less effective primary care than suburban children. 3) Mandatory hospitalizations will follow the same socioeconomic gradient followed by discretionary admissions. Hospitalization rates for life-threatening conditions approximate true incidence rates. The majority of mandatory admissions are attributable to severe trauma or serious bacterial infections, conditions more common among impoverished children. 4) Hospitalization rates for the unlikely class will vary little among socioeconomic areas. Because conditions in this class are almost always managed safely without hospitalization, the decision to hospitalize involves little uncertainty and little variation was anticipated.
Agreement on Classification of Diagnoses
Final classifications (based on ratings of the expert panel) and preliminary classifications (performed by the authors) were identical for 77.9% of diagnoses presented to the panel. Agreement beyond that expected by chance (kappa) was .61 (95% CI: .47 to .74), a level that is considered substantial.26,27 Agreement among individual panel members was moderate. Values of kappa comparing each rater with all other raters on the 95 rated diagnoses averaged .50 (95% CI, .37 to 63).
Attributes of Socioeconomic Areas
As indicated in Table 1, striking differences exist in many sociodemographic attributes between the inner city and the suburbs, as well as distinct, stepwise gradients in these attributes in moving from inner city, to other urban areas, and to the suburbs. Racial and economic segregation is particularly striking. African-Americans accounted for 61.8% of inner-city births versus less than 2.6% in the suburbs. Medicaid covered 65.0% of inner-city births compared to only 5.7% in the suburbs. Distinct area gradients were also noted for timeliness of prenatal care, low birth weight, maternal smoking, ethnicity, educational achievement, adolescent pregnancy rates, housing units without phones and housing units without automobiles, proportion of preschool children living in poverty, unemployment rates, and per capita income. Moderate changes in socioeconomic indicators over the 7-year study period were noted. This reflected increases in minority populations and poverty that were concentrated in the city, not changes in Medicaid eligibility. Differences among the three socioeconomic areas, however, remained distinct. The proportion of all hospitalizations covered by Medicaid for the first 2 (1985 to 1986), middle 3 (1987 to 1989), and last 2 (1990 to 1991) study years, for example, were as follows: inner city: 69.1%, 73.1%, and 77.8%; other urban: 43.8%, 42.8%, and 54.5%; and suburbs: 9.9%, 12.5%, and 15.9%. Similarly, the proportions of births accounted for by African-Americans for the same periods were: inner city: 51.9%, 57.0%, and 59.5%; other urban: 12.9%, 15.6%, and 17.8%; and suburbs: 1.7%, 2.6%, and 2.9%.
Attributes for the 2111 child-years that could not be attributed to a specific socioeconomic area because of missing data were very similar to those for all births, indicating that lack of information about socioeconomic area for these infants was unlikely to bias results.
Two hospitals accounted for 94.5% of the 7883 hospital admissions occurring over the 7-year observation period. Overall hospitalization rate was 50.3 per 1000 child-years. The distribution of hospitalizations by decision class for all areas combined and for different socioeconomic areas is shown in Table 3. The discretionary category was the largest group, accounting for 59.1% of admissions, followed by mandatory, and sometime hospitalizations, accounting for 17.7 and 15.2%, respectively. Discretionary surgery accounted for 6.2% of admissions, and unlikely accounted for only 1.8%.
Variations in Hospitalization Among Socioeconomic Areas
Rates for all hospitalizations in each socioeconomic area followed a steep gradient with rates for inner-city children of 82.9 per 1000 compared to rates of 51.3 and 38.1 for other urban and suburban children. In calculating odds ratios (OR), odds for hospitalization of suburban infants were used as the base odds. The odds of hospitalization for inner-city children were more than twice that for suburban children (OR, 2.28; 95% CI, 2.16 to 2.41, P< .001), and the odds of hospitalization for other urban children were also greater than for suburban children (OR, 1.37; 95% CI, 1.29 to 1.45).
The largest socioeconomic gradient was observed for the discretionary class, with an OR for the inner city of 2.88 (CI, 2.69 to 3.08; P < .001), and an OR for other urban of 1.55 (CI, 1.44 to 1.67; P < .001). A substantial gradient was also observed for the mandatory class with an OR for inner-city hospitalization of 2.20 (CI, 1.94 to 2.49; P < .001) and an OR for other urban hospitalization of 1.36 (CI, 1.19 to 1.55;P < .001). Although the rate was much lower, the gradient for the unlikely class was similar to that for mandatory admissions. For unlikely admissions, the OR for the inner city was 2.57 (CI, 1.73 to 3.83; P < .001), and an OR for other urban was 1.69 (CI, 1.12 to 2.55; P < .05). Rates for the sometime class did not vary significantly among socioeconomic areas. Discretionary surgery was slightly greater for inner-city infants than suburban infants with an OR of 1.62 (1.31 to 2.00;P < .001), but there was no difference in discretionary surgery between other urban and suburban infants (OR, 1.18; 95% CI, .95 to 1.46).
Predicting Variation Among Zip Code Areas
For the 37 zip code areas, correlations were high between discretionary hospitalization rates and the variables included in Table1 reflecting sociodemographics, housing, late prenatal care, and health related behaviors, ranging from .75 for maternal smoking to .95 for maternal education. All these correlations were significant at least at the .001 level (two-tailed). All correlations between mandatory hospitalization rates and these variables were also significant at least at the .001 level, but these relationships were not as strong; they ranged from .61 for maternal smoking to .80 for initiation of prenatal care after the sixth month of pregnancy (late prenatal care). In contrast, none of the correlations between sociodemographic factors and the classes sometime or discretionary surgery were statistically significant. Unlikely hospitalization rates were not considered in correlation or regression analysis because they accounted for a very small proportion of admissions.
Relationships with hospitalization rates were assessed further in multiple regression analysis (Table 4). In stepwise analysis, only one variable entered the regression model predicting discretionary hospitalization rates—proportion of mothers in the zip code area with less than a high school education. Education level was a highly significant predictor (P < .0001), and this variable alone, explained 89% of the variance in the rate for discretionary admissions. Only late prenatal care entered the regression model in stepwise analysis predicting mandatory hospitalization rates. A highly significant predictor (P < .0001), late prenatal care explained 65% of the variance. None of the social or demographic characteristics of the 37 zip code areas predicted sometime or discretionary surgery hospitalization rates in regression analysis.
Further analysis of zip code areas supported the validity of using Medicaid births as a primary criterion for classifying zip code areas into socioeconomic areas. Moreover, this analysis supported the hypothesis that socioeconomic characteristics of areas (aggregate level data) generally applied to individual children hospitalized from these areas (individual level data). The median and range for the proportions of Medicaid births for the 4 inner city, 7 other urban, and 26 suburban zip code areas were 66.4% (58.5 to 77.1), 29.0% (17.0 to 40.4), and 5.3% (1.3 to 16.3), respectively. The correlation between the proportions of births covered by Medicaid and the proportions of hospitalizations covered by Medicaid for the 37 zip code areas was high (.88), indicating that the proportion of Medicaid births for these geographic areas is highly representative of the Medicaid status of children that are hospitalized. Consistent with this finding, the proportions of Medicaid covered births and hospitalizations, respectively, for the inner city, other urban, and suburban socioeconomic areas were 65.0 and 73.7%, 26.8 and 47.1%, and 5.7 and 12.8%.
Patient-Class Hospitalized Rates
A total of 4923 children accounted for the 7883 hospital admissions and for 6915 patient-class hospitalizations. Thus, children hospitalized averaged 1.60 hospitalizations per child overall and 1.40 hospitalizations per child in a unique patient class. Patient-class hospitalized rates were somewhat lower than hospitalization rates, but they followed a similar pattern and had similarly narrow 95% CIs as the hospitalization rates presented in Table 3. Overall patient-class hospitalized rate was 44.2 (43.9 to 44.4) per thousand child-years and varied from 33.3 (33.1 to 33.7) for the suburbs to 45.6 (45.1 to 46.1), and 71.9 (71.4 to 46.1) in other urban and inner-city areas, respectively. Patient-class hospitalized rates for the mandatory class were 6.3 (6.2 to 6.5), 8.5 (8.2 to 8.7), and 14.0 (13.6 to 14.3) in these three areas, and they were 17.4 (17.1 to 17.6), 26.3 (25.9 to 26.7), and 45.7 (45.2 to 46.3) for the discretionary class.
Reliability and Validity of the Classification Scheme
Panel members demonstrated a high level of agreement in classification of diagnoses, and agreement between panel classifications and preliminary ratings by authors was also high. Content and face validity of the classification (Table 2) has been discussed. Empirical construct validation is, fundamentally, an ongoing process of hypothesis testing in which each supportive finding strengthens the network of evidence that is required to establish validity firmly.28-30 Existing empirical evidence for validity (see Methods) is enhanced by findings of this study. Distributions of discretionary surgery and sometime classes, the mandatory class, and the discretionary class were consistent with hypotheses 1 through 3, respectively.
Contrary to hypothesis 4, variation for unlikely admissions was substantial. Variation was not expected because there is little uncertainty about the outcome of the vast majority of illness episodes with ICD codes that fall in this category. We speculate that observations are attributable to the fact that ICD coding represented discharge diagnoses, not admission diagnoses. We expect that the small number of unlikely admissions were mostly admitted because of fever without a source, and that the labels used for discharge diagnoses (eg, acute otitis media, upper respiratory tract infection) were selected only after more serious etiologies were ruled out.
Limitations of cross-sectional and, in particular, ecologic study designs should be recognized. Causal relationships between aggregated characteristics of a geographic area and events occurring to individuals dwelling in the area cannot be inferred. Nevertheless, analysis of the 37 zip code areas was consistent with the hypothesis that socioeconomic characteristics of areas (eg, aggregate level data on Medicaid births) generally applied to individual children hospitalized from these areas (eg, individual level data on Medicaid coverage for infants hospitalized). Such evidence of homogeneity within areas suggests that ecologic fallacy is much less likely.
Analysis focused on hospitalization rates rather than rates of children hospitalized (eg, patient-class hospitalized rate) because each hospitalization represents a large expense and each hospitalization may be viewed as a unique product of a geographic area and its environment. Hospital admission for an individual child, however, is associated with increased likelihood of subsequent hospitalization. Clustering of events such as hospitalization violates the assumption of independence and yields standard errors that underestimate confidence intervals for point estimates of hospitalization rates. Comparison of patient-class hospitalized rates with hospitalization rates allows one to assess the importance of patient clustering of hospitalizations on estimates. Standard errors of the patient-class hospitalization rates may be viewed as an upper bound for true standard errors of hospitalization rates. Although rehospitalization was relatively common, differences in rates among the three socioeconomic areas were large and many fold greater than the standard errors, whether hospitalization rates or patient-class hospitalized rates are considered.
Regression analysis should be interpreted cautiously. With only 37 observations (zip code areas) in this analysis and the high colinearity among sociodemographic variables, the relationship between mothers' educational levels and discretionary hospitalization rates can only be inferred to indicate a strong relationship with social factors, not with educational level, per se. The lack of any relationship in correlation and regression analysis between social factors and rates for sometime and discretionary surgery admissions supports the validity of the classification, as does the presence of strong relationships between social factors and both mandatory and discretionary hospitalization rates.
Variation in Overall Hospitalization Rates
Findings of this study highlight the importance of socioeconomic status to child health and to the cost of child health care. A large portion of the difference in per capita health care costs between inner-city and suburban infants appear to be attributable to hospitalization rates that were more than two-fold greater for inner-city areas. Plausible hypotheses explaining this difference fall in two categories, morbidity burden hypotheses and health system hypotheses. Morbidity burden hypotheses, explaining variation in hospitalization rates based on socioeconomic differences in morbidity rates or case severity, are consistent with the strong association between socioeconomic status and childhood morbidity burden that has long been recognized. Health system hypotheses attribute variation to characteristics of the health care system, provider attributes, parent behavior, or systematic differences in the interactions between the health care system and parent behavior.
Variation in Mandatory Admissions
Morbidity burden hypotheses are supported by findings relating to mandatory hospitalizations. Hospitalization rates for conditions included in the mandatory category should closely approximate their incidence rate in an area such as Monroe County where hospital care for severe illness is uniformly available across socioeconomic groups. The odds of a mandatory admission for an inner-city infant was 2.4-fold greater than for a suburban infant, with a step-wise gradient in mandatory admission rate from suburban to other urban and inner-city areas. Potential reasons for higher rates of severe illness among lower socioeconomic infants include constitutional and environmental differences. Environmental differences play a substantial role in the etiology or exacerbation of many conditions in this class (eg, severe trauma, severe bacterial infections). In contrast, it appears unlikely that constitutional differences are a major determinant of differences between socioeconomic areas. This is supported by observations that rates for discretionary surgery and sometime admissions, classes that contain many congenital conditions (Table 2), are similar. Moreover, genetic factors appear to play a role in very few conditions in the mandatory class (Table 2).
The possibility that health system hypotheses explain a large portion of variation among mandatory admissions appears less likely. Variation in immunization rates for Haemophilus influenzae type b vaccine might explain some socioeconomic variation at present, but this probably was a less important factor during the observation period because vaccination was not recommended for use in infants beginning at 2 months of age until October, 1990.31 Incidence of severe head trauma, severe burns, or drowning might be influenced by safety counselling, but it appears unlikely that the impact of this intervention would be sufficient for differences in access or quality of health care to account for much of the observed socioeconomic variation in mandatory admissions.
Variation in Discretionary Admissions—Environmental Determinants
Because effects of constitutional differences on variations are doubtful, and interventions to prevent constitutional differences are generally unknown, environmental effects on morbidity burden are a more immediate concern. Social and physical environments vary widely across socioeconomic areas, and a large body of evidence supports the view that they are important determinants of morbidity in the discretionary class. Environment plays a substantial role in the etiology or exacerbation of asthma,32-36bronchiolitis,37 gastroenteritis,38-40 soft tissue infections,41 lead exposure, and failure to thrive.42 Evidence supports several mechanisms for increased morbidity burden among lower socioeconomic infants due to these conditions. Mechanisms involve increased exposure to environmental tobacco smoke, household crowding, older siblings with viral illness, allergens, household hazards, lead paint, suboptimal caretaking, parental depression, and accidental injury.
Variation in Discretionary Admissions—Health System Determinants
We speculate that socioeconomic disparities in hospitalization rates for discretionary conditions are attributable to differences both in morbidity burden and in the health care system, and that health system effects are at least as important. In part, this view is based on the degree of variation for discretionary conditions (inner city:suburban OR of 2.9) which is large compared to effect sizes typical for environmental exposures. This speculation is also based on our clinical experience and prior studies addressing variation in hospitalization for discretionary conditions.
Many studies demonstrate wide variation among similar socioeconomic areas,4-6,8 supporting the belief that, for diagnoses included in the discretionary category, the decision to hospitalize is often truly discretionary. Discretionary admissions for infants involve high-stakes decisions. False positives (unnecessary hospitalizations) incur high costs, but potential consequences of false negatives (infant death in a child unsafely sent home from the emergency department) are disastrous and unacceptable.
Many plausible mechanisms underlie health system hypotheses. Variation in area-specific practice patterns within this single medical community may be less important than variation among different medical communities. As previously noted, two hospitals accounted for 94.5% of all pediatric admissions in this study. Among discretionary hospitalizations, 94.8% passed through the emergency departments of these two hospitals, which are staffed by a single group of pediatric residents and a closely-linked group of emergency department pediatric attending physicians. Also, hospital bed availability is not a determinant because these two hospitals provide, in effect, a single community bed supply, and they serve similar, diverse population groups.
On the other hand, providers based in the Rochester suburbs may interact quite differently with families of ill infants than providers for inner-city infants. System hypotheses probably relevant to differences in discretionary admissions in this study encompass individual provider factors (eg, medical knowledge and its application, communication skills, comfort with uncertainty, and risk aversion), structural health system attributes (eg, telephone coverage, condition specific hospitalization criteria, alternatives to hospitalization for acute illness care, and transportation systems), and family characteristics that affect interactions with the health care system and ability to care for an acutely ill child safely at home (eg, observation and communication skills, telephone and automobile availability). The observation that maternal education was the strongest predictor of discretionary hospitalization and explained 81% of its variance is consistent with the possibility that the family-health system interface was a key determinant of variation. Findings in a recent study comparing care for children hospitalized for asthma in Boston, New Haven, and Rochester suggest that higher severity thresholds for admission explain much of the observed differences.43 Differences in severity threshold might also explain differences in hospitalizations among socioeconomic areas within the Rochester area, a hypothesis that warrants study.
Implications for Evaluation of Health Services
Variation in hospitalization rates for ambulatory care sensitive conditions has previously been recognized as a key indicator of health care access or quality.10-14 Hospitalization for conditions in the discretionary class may be sensitive to the quality of primary care for several reasons, as follows: 1) Early intervention may prevent progression to a level of physiologic derangement requiring hospitalization. Antibiotic therapy early in the course of cellulitis, oral electrolyte therapy early in the course of gastroenteritis, and medications early in the course of an asthmatic attack are common examples. Early implementation of these interventions may require ready access to a telephone coverage system staffed by providers that communicate well. In addition, early intervention may require readily accessible ambulatory care facilities. 2) More effective illness management may also reduce hospitalization. Effective management requires appropriate application of medical knowledge. Examples include appropriate use of antibiotics, asthma monitoring techniques and asthma medications. 3) Better decision making may also avoid more hospitalizations. To reduce hospitalization, decision patterns with greater specificity must evolve (patterns in which a larger proportion of children who do not need hospital care are sent home from the emergency department). Simultaneously, high sensitivity must be maintained (all illness episodes that require hospitalization must be admitted). Provider function and health services structure might influence decision making because of variability in psychologic attributes of providers (eg, comfort with uncertainty, risk aversion, and need for control), and provider ability to communicate well. Attributes of the health care system affecting decision making may include structural elements that affect liability. Ambulatory facilities that allow extended visits involving relatively intensive treatment may allow more effective care and better decision making. Examples include intravenous hydration, repeated nebulizer treatments, and prolonged observation following head trauma over a 4- to 6-hour period. 4) Finally, systems that facilitate careful follow-up may enable better hospitalization decision making because they allow primary care providers and families to be confident that care at home is safe since return to a medical facility will rapidly follow if deterioration occurs.
Timely intervention, effective illness management, accurate hospitalization decision making and careful follow-up are congruent with cardinal objectives of primary care.44 Accordingly, hospitalization rates for conditions in the discretionary class might be considered as a measure for the effectiveness of primary care. In comparisons where similarity in morbidity burden may be established or reasonably assumed, discretionary hospitalization rates may provide a useful measure for the effectiveness of primary care. Thus, discretionary hospitalization rates might be useful in monitoring changes in primary care over time within a single, socioeconomically stable area. They might also be useful in cross-sectional comparisons of health systems among different geographic areas that are socioeconomically similar. Use of discretionary hospitalization rates to compare health systems among areas that are socioeconomically different is more complex and would require valid health status measurements to allow adjustment for differences in morbidity burden. Further research focused on the relationship between effectiveness of primary care and hospitalization for discretionary conditions, particularly individual level analyses, appears warranted.
Implications for Developing Health Services
Although the limitations of existing measures of health status make it impossible to determine the relative importance of morbidity burden versus system etiologies of socioeconomic differences in hospitalization rates, there are several reasons to enhance services for inner-city children now. 1) Elimination of the system component of the suburban-inner city difference might, in itself, result in major cost reduction. 2) Elimination of the system component might be accomplished more readily than the environmental changes that might be required to eliminate the morbidity burden component. 3) Inpatient care is so costly that even relatively small reductions might have a financial impact that far exceeds the cost of interventions required to produce them. 4) Ultimately, which hospitalizations can be avoided and what proportion of the socioeconomic disparity in hospitalization rates can be eliminated will only be determined by efforts to intervene.
Eliminating the health system component of this costly socioeconomic disparity represents a challenge that may require major innovations in health services delivery. Many inner-city families bring limited resources to the family-health system interface. Inner-city primary care providers may need to devote more time and effort to communication and issues related to cultural diversity than those in the suburbs. Inner-city facilities may require more elaborate and more expensive medical information and communications systems. More outreach and more intensive follow-up efforts in the inner city, perhaps including extensive use of home-visiting nurses, might also contribute to elimination of the system component. Such innovations may be important to the success of managed-care programs that are now addressing the needs of the increasingly large number of children and families dwelling in poverty.
This study was supported in part by grant MCJ-360571 from the Maternal and Child Health Research Program (Title V, Social Security Act), Health Resources and Services Administration, Department of Health and Human Services.
The authors gratefully acknowledge the assistance of the Jeffry Mailloux, Jill Szydlowski, and Beverley Voos of the Rochester Healthcare Information Group for provision of hospitalization data; and Robert Byrd, MD; Charles Callahan, MD; Gregory Conners, MD; Lynn Garfunkel, MD; Viking Hedberg, MD; King Hom, MD; Jonathon Klein, MD; Marc Lampell, MD; Stanley Schaeffer, MD; and Peter Szilagyi, MD, for participation as members of the expert panel that classified diagnoses.
- Received May 9, 1995.
- Accepted October 11, 1996.
Presented in part at the 34th Annual Meeting of the Ambulatory Pediatric Association, Seattle, Washington, May 4, 1994.
- DRG =
- diagnosis-related group •
- ICD =
- International Classification of Diseases •
- CI =
- confidence interval •
- OR =
- odds ratio
- ↵Hahn B, Lefkowitz D. Annual Expenses and Sources of Payment for Health Care Services. Rockville, MD: Agency for Health Care Policy and Research; US Department of Health and Human Services; November 1992, Pub. No. 93–0007
- ↵Lewit EM, Monheit AC. Expenditures on health care for children and pregnant women. The Future of Children. Winter 1992;(Winter):95–114
- Goodman DC,
- Fisher ES,
- Gittelsohn A,
- Chang CH,
- Fleming C
- ↵Wennberg J, Kim SYS. Common uses of hospitals: a look at Vermont. In: Harvard Child Health Project, II. Developing a Better Health System for Children. Cambridge, MA: Ballinger; 1977:353–382
- Soulen JL,
- Duggan AK,
- DeAngelis CD
- Conway SP,
- Phillips RR,
- Panday S
- Billings J,
- Zeitel L,
- Lukomnik J,
- Carey TS,
- Blank AE,
- Newman L
- ↵Weiss KB, Budetti P. Examining issues in health care delivery for asthma: background and workshop overview. Med Care. 1993;31:MS9-MS19
- ↵Center for Health Economics Research, for The Robert Wood Johnson Foundation. Access to Care: Key Indicators for Policy. Princeton, NJ: Robert Wood Johnson Foundation; 1993:56–57
- Wennberg JE,
- Barnes BA,
- Zubkoff M
- Paul-Shaheen P,
- Clark JD,
- Williams D
- Hill AM
- ↵International Classification of Diseases, 9th Revision: Clinical Modification. Washington, DC: US Department of Health and Human Services; 1980. DHHS Pub No. (PHS) 80–1260
- ↵Feinstein AR. Clinical Epidemiology: The Architecture of Clinical Research. Philadelphia, PA: WB Saunders; 1985:186
- ↵Streiner DL, Norman GR. Health Measurement Scales: A Practical Guide to Their Development and Use. New York, NY: Oxford University Press; 1989
- American Academy of Pediatrics, Committee on Infectious Diseases
- Stanhope JM,
- Rees RO,
- Mangan AJ
- Lebowitz MD
- ↵Spence J, Walton WS, Miller FJW, Court SDM. A Thousand Families in Newcastle-Upon Tyne. London, England: Oxford University Press; 1954
- Newberger EH,
- Reed RB,
- Daniel JH,
- et al.
- Homer CJ,
- Szilagyi P,
- Rodewald L,
- Bloom SR,
- Greenspan P,
- Yazdgerdi S,
- Leventhal JM,
- Finkelstein D,
- Perrin JM
- ↵Charney E, Alpert JJ. The Education of Physicians for Primary Care. Washington, DC: US Department of Health, Education, and Welfare, Bureau of Health Services Research; DHEW Pub. No. (HRA)74–3113
- Copyright © 1997 American Academy of Pediatrics