Community-Level Correlates of Hospitalizations for Persons With Schizophrenia

To conduct the analysis, it was necessary to define communities geographically, and we chose the county as the unit of analysis. Our choice was purely pragmatic and was motivated by the fact that a great deal of information about communities is aggregated at the county level. Counties offer an appropriate level of aggregation for analyzing hospitalizations, whereas a smaller geographical unit (such as zip codes) would have yielded zero hospitalizations too frequently to be analyzed.

Schizophrenia-related hospitalizations were extracted from the Statewide Inpatient Database developed by the Agency for Healthcare Research and Quality. The database contains the universe of hospital discharge records from all hospitals in participating states. Data from 2000 were extracted for the following 14 states: Arizona, Colorado, Florida, Iowa, Kentucky, Maine, New Jersey, New York, North Carolina, Oregon, South Carolina, Washington, West Virginia, and Wisconsin. Although 17 states contributed data to this database in 2000, only 14 provided patient zip code data needed for this project. Note that the counties included in the analysis were not randomly sampled. However, there were no substantial differences between the counties in the analytical data set and all U.S. counties with respect to the community-level correlates examined.

The database includes data about the patient's zip code (regardless of residential location), primary diagnosis, and up to nine secondary diagnoses. Because unique patient identifiers were not available from most states, the data reflect multiple admissions for some individuals. Hospitalizations with a primary or secondary diagnosis of schizophrenia ( ICD-9-CM code 295.xx) were geocoded to the patient's county of residence. Using age-sex categories defined by the U.S. Census Bureau, we categorized each inpatient into one of 20 age-sex groups (age groups: 0–24, 25–29, 30–34, 35–44, 45–54, 55–59, 60–64, 65–74, 75–84, and 85 and older). The hospitalization rate for each age-sex category was calculated by dividing the number of schizophrenia-related hospitalizations in each age-sex category aggregated across the 14 states by the population in each age-sex category aggregated across the 14 states as reported in the 2000 census. The expected number of schizophrenia-related hospitalizations for each county was calculated by multiplying the age-sex hospitalization rates by the county population in each age-sex category and summing these products across all age-sex categories. The indirectly standardized hospitalization rate (SHR) for each county was then defined as the ratio of the observed to the expected number of hospitalizations ( 22 ).

Rurality in 2000 was measured at the county level with two indicators. The first was that the county did not meet the U.S. Office of Management and Budget's definition of a metropolitan statistical area. The Office of Management and Budget defines a county as a metropolitan statistical area if it contains an urbanized area (population ≥50,000) or is adjacent to such a county and 25% of the employed population commutes to the urbanized area (or vice versa). Rurality was also measured at the county level with urban influence codes (UICs) ( 23 ). For this measure counties are divided into 12 categories according to their population, adjacency to an urbanized area, and the proportion of workers commuting to urbanized areas.

Sociodemographic variables included percentage of the population that was African American, Asian American, Native American, or Hispanic as defined and measured by the 2000 census. Other socioeconomic variables from the census included percentage below the federally designated poverty level, percentage with a high school education, median household income (1999 dollars), and percentage unemployed (among persons aged 16 and older). With 2000 data compiled by the U.S. Department of Agriculture, counties were classified as housing stressed if 30% or more of households had one or more of the following: lacked complete plumbing, lacked a complete kitchen, renter or owner paid 30% or more of income for mortgage or rent, or had more than one person per room. Economic data for 2000 generated by the U.S. Department of Agriculture classified counties into one of six mutually exclusive categories indicating whether the economy of the county was dependent on farming, mining, manufacturing, federal or state government, or services or was not dependent on a specialized sector ( 24 ).

Health system data for 2000 from the Health Resources and Services Administration's Area Resource File included the number of nonpsychiatrist physicians, psychiatrists, psychologists, and social workers per 1,000 people, number of hospital beds per 1,000 people, the market penetration rate of HMOs, and dummy variables indicating whether the county had a community mental health center or was designated as a health professional shortage area ( 25 ). Finally, we also included the longitude and latitude of the counties, because hospitalization rates seem to vary east to west and north to south.

Counties were specified as the unit of the ecological regression analysis. On the basis of previous research ( 23, 24 ), we expected SHRs among contiguous (neighboring) counties to be correlated and the observations to not be independently distributed ( 26 ). Therefore, we used spatial regression analysis, which accounts for the spatial autocorrelation among contiguous counties ( 27, 28 ). Because the spatial regression models used in the second stage of the analysis were computationally intensive, it was necessary to first identify a parsimonious model specification with a nonspatial regression analysis. Backward selection techniques (using negative binomial regression and a cutoff level of p<.2) were used to identify variables to include in the parsimonious model specification. All community-level correlates described above were included as candidates in the backward selection analysis.

The spatial regression analysis used a conditional autoregressive model to account for the spatial autocorrelation among contiguous counties. Because some counties with small populations had extremely high or low SHRs (probably because of random variation over time) ( 29 ), we used Bayesian smoothing methods, which took advantage of the spatial autocorrelation present in the data ( 30 ). Specifically, we specified the Bayesian-Poisson conditional autoregressive model proposed by Besag and colleagues ( 31 ), as implemented in the statistical software GeoBugs ( 32 ). This Bayesian spatial regression model is described in detail in our companion article, which describes the model's use in analyzing depression-related hospitalization rates ( 27 ), and by Moreno and colleagues ( 33 ), who used it to identify spatial clustering of patients with schizophrenia in Spain.

We also estimated the parameters of the Bayesian-Poisson model without the conditional autoregressive term to determine whether spatial autocorrelation was present in the data. Specifically, we used the deviance information criterion to compare whether the Bayesian-Poisson autoregressive model had better goodness of fit than the simple Bayesian-Poisson model ( 34 ). Smaller deviance criterion values represent better goodness of fit. However, because the criterion has no known distribution, a statistical test was not performed.

We identified outlier counties that had SHRs 25% less than and 33% greater than the standardized rate (or SHRs <.75 and >1.333). Note that these cutoffs are symmetrical around an SHR value of 1: 1.33=1/(1–.25). Two methods were used to identify outlier counties. The first method examined raw SHR data. However, because the SHRs in counties with small populations are sensitive to random variations in hospitalizations over time, this method was prone to incorrectly identifying small counties as outliers. The population of the smallest county in the sample was 436. Therefore, the second method examined smoothed SHR data, which exploited the spatial autocorrelation. Specifically, smoothed SHRs were calculated as the ratio of the smoothed observed number of hospitalizations to the smoothed expected number of hospitalizations. The smoothed observed number of hospitalizations was a conditional expectation derived from the Bayesian-Poisson conditional autoregressive model that included an intercept, a spatially correlated error term for each county, and a subject heterogeneity error but did not include any covariates. Inclusion of the spatially correlated error term smoothed the observed number of hospitalizations, especially for the counties with small populations. Inclusion of the subject heterogeneity error took into account the variability across counties. The smoothed expected number of hospitalizations was a conditional expectation derived from the Bayesian-Poisson conditional autoregressive model that included an intercept and a spatially correlated error term for each county but did not include a subject heterogeneity error or any covariates. Counties that had smoothed SHRs 25% less than and 33% greater than the standardized rate (smoothed SHRs <.75 and >1.333) were identified as outliers.