All hospitalizations of those aged 65 years or older were considered (N = 167,793). Hospitalizations with Department of Defense (DOD) as payer were excluded (n = 7,987) because the DOD does not consistently report race/ethnicity data. Hospitalizations without valid race/ethnicity data were excluded (n = 3,073) as were those not reporting a race/ethnicity for the 5 largest ethnic subgroups in Hawai‘i (ie, Japanese, Chinese, Native Hawaiians, Filipinos, or white), which were the focus of this study. To be consistent with population and disease prevalence totals used for rate denominators (estimated for Hawai‘i residents), individuals not from Hawai‘i were excluded (n = 5,362). We also excluded transfers and unknown admission source (n = 4,517) to meet the definitions of the Agency for Healthcare Research and Quality (AHRQ) for diabetes-related preventable hospitalizations.

HHIC data include a master patient identification variable that tracks individuals across all hospitals in the state. Considering unique individuals confirms that multiple visits by members of certain racial/ethnic groups are not driving health disparities, an important issue in diabetes where racial disparities are seen in readmissions (18). After exclusions, there were a total of 146,854 eligible hospitalizations from 72,876 unique patients.

We followed AHRQ definitions to identify DRPH (19). Specifically, we included 1) uncontrolled diabetes without mention of a short-term or long-term complication, 2) diabetes with short-term complications (eg, ketoacidosis, hyperosmolarity, coma), 3) diabetes with long-term complications (eg, renal, eye, neurologic, circulatory, or complications not otherwise specified), and 4) lower extremity diabetes-related amputations. More detail is available from AHRQ (www.qualityindicators.ahrq.gov/Downloads/Modules/PQI/V31/pqi_guide_v31.pdf).

The HHIC race/ethnicity variable was created from race/ethnicity categories available consistently across all hospitals in Hawai‘i from December 2006 through December 2010 (17). Race/ethnicity data are typically provided by patient self-report at intake and include only 1 primary race. Mixed-race individuals are represented as their primary race of identification or excluded by being part of the “other” racial/ethnic category.

In multivariable models, we included sex, because diabetes prevalence varies by this factor (20); comorbidity, defined by the Charlson comorbidity index (CCI) (21); payer (Medicare, Medicaid, private, and other); location of residence (lives on Oahu vs another Hawaiian island), based on evidence that access to care is often worse on the other islands compared with the more urban Oahu; and age (continuous), because of differences in age structure by racial/ethnic groups. In summary tables, unadjusted and adjusted rates are presented by sex to be more useful for clinical and policy purposes.

When calculating rates of preventable hospitalizations, population totals within the relevant geographic region (eg, the state) are often used as denominators (22). This takes into account the fact that subpopulations have different sizes, allowing identification of the burden of preventable hospitalizations in specific subpopulations. Recent research has emphasized that the disease burden by subpopulations should also be considered in preventable hospitalizations rates (22), particularly when there are known disparities in disease prevalence across subgroups (23). To understand the full burden of DRPH specifically among each racial/ethnic subgroup in Hawai‘i, we calculated DRPH rates first using population totals, and then using disease prevalence totals.

Population totals by sex and race/ethnicity combinations as well as the number of people with diabetes by subgroup were obtained from 2007–2010 Hawai‘i Behavioral Risk Factor Surveillance System (BRFSS) data. The BRFSS is the standard for state-level diabetes rates in Hawai‘i. Self-reported diabetes prevalence was calculated yearly from this representative statewide survey using standard methods (24). Four years of survey data were combined to provide more reliable state-level estimates of disease burden by race/ethnicity. The BRFSS race/ethnicity classification scheme is compatible with that used by HHIC because it is self-reported and includes racial/ethnic categories for the 5 major racial/ethnic groups in Hawai‘i.

Characteristics of the patients with a DRPH were summarized by descriptive statistics for each racial/ethnic subgroup and compared among subgroups using χ² tests or Fisher’s exact tests (for categorical variables) and analysis of variance (ANOVA) or nonparametric Kruskal–Wallis test (for continuous variables). For patients with multiple visits, we used the patient’s first hospitalization in the analysis.

The unadjusted average annual rates of DRPH by patient among AA/PI subgroups and whites were calculated by sex first by using BRFSS population totals and then using population-level totals of diabetes prevalence as denominators. Unadjusted rate ratios (RR) of DRPH by patient were then calculated by dividing the unadjusted rate for each racial/ethnic group by the unadjusted rate for whites. A possible disparity for an AA/PI subgroup relative to whites is represented as RR greater than 1.0. Next, multivariable models were developed to estimate diabetes-related potentially preventable hospitalization rates by patient adjusting not only for sex and race/ethnicity but also for other explanatory factors that may predict hospitalization (ie, comorbidity, residence in Oahu, age, and insurer). To fully understand the portrait of DRPH by racial group, we used 2 multivariable models. In the first model (model A), adjusted rates were calculated by racial/ethnic group by using population totals for rate denominators. This represents the most common portrait of potentially preventable hospitalizations and matches AHRQ guidelines. In the second model (model B), adjusted rates were calculated by racial/ethnic group by using diabetes prevalence for the rate denominators. This represents a portrait of preventable hospitalizations that may more fully capture disparities and shed more light on access to care or quality of outpatient care issues.

In both models, percentage of public insurance (Medicare and Medicaid) and median CCI for each race and sex combination were calculated considering insurer and comorbidity factors, respectively, and entered into models. Overall hospitalization data were modeled using negative binomial regressions models adjusting for possible overdispersion (25). Multivariable adjusted rate ratios (aRR) of patients with DRPH by patient were derived from the final models for each racial/ethnic group compared with whites. All data analyses were performed in SAS 9.3 (SAS Institute Inc, Cary, North Carolina). A 2-tailed P value of less than .05 was regarded as significant.

In unadjusted models, with population totals as the rate denominator, disparities in DRPH were seen for women and men in all AA/PI racial/ethnic groups compared with whites, with RRs ranging from 1.32 in Chinese men to 3.98 in Filipino women (Table 2). These findings remained when other factors were controlled. Even in the adjusted model (model A), a significantly higher population-level burden of DRPH was seen for women and men in all nonwhite racial/ethnic groups compared with whites. Adjusted RRs ranged from 1.94 (95% CI, 1.30–2.91) in Chinese women to 4.18 (95% CI, 3.31–5.27) in Filipino women.

When the rates of DRPH by patient only among those with diabetes were considered, for some AA/PI groups, particularly women, the higher population-level burden of DRPH was explained by the higher diabetes burden in AA/PI groups compared with whites. When the number of people with diabetes was used as the denominator, even in unadjusted models, no disparity in DRPH was seen for Chinese women (RR = 0.90) and men (RR = 0.91) and Japanese women (RR = 0.95) relative to whites. Whites appear to have a disparity relative to these AA/PI groups (Table 3).

However, this was not true for all AA/PI groups. Even when the higher prevalence of diabetes was considered, disparities in DRPH remained for Filipinos, Native Hawaiians, and Japanese in unadjusted models. In fully adjusted models, disparities remained for Filipino women (aRR = 1.61; 95% CI, 1.28–2.03) and Filipino men (aRR = 2.26; 95% CI, 1.71–2.99), Native Hawaiian men (aRR = 1.59; 95% CI, 1.11–2.28), and Japanese men (aRR = 1.86; 95% CI, 1.04–3.33) compared with whites.

Individuals had to have a diabetes diagnosis to meet the AHRQ definition for DRPH. Because 40% of diabetes in the United States is undiagnosed (27), some visits that should be classified as a DRPH may not be. Also, the self-reported diabetes prevalence obtained from the BRFSS likely underestimates true prevalence, both because not all people with diabetes are aware of their condition (which may vary by race/ethnicity) and because the BRFSS sample excludes people without a telephone, which may include people particularly likely to be hospitalized for diabetes (28). If whites and all AA/PI groups are similarly underestimated, the underestimation would affect point estimates, but not relative rates.

Although we have a comprehensive, state-level data set, we include only 1 state, and it may not be representative of other areas. However, because access to culturally appropriate care may be worse for AA/PI groups in many other settings, our study may actually underestimates AA/PI disparities.

Our analyses are based on administrative data, which have some general limitations (29) and lack some variables that would have been useful, such as demographics (eg, education), modifiable risk factors (eg, obesity), differences in access (eg, having a primary care provider), pathways to care (eg, emergency department vs outpatient setting), or quality of care. We also lack information about the time since a diabetes diagnosis. These are all areas for future research. Also, although the HHIC race/ethnicity data have been verified to be collected consistently across hospitals, this is not guaranteed in all cases. However, in this regard we share these limitations with many studies on this topic because the AHRQ metrics are designed to use administrative data.

We identified 1,815 hospitalizations by 1,515 elderly patients in a 4-year period in Hawai‘i that were potentially avoidable with better primary care for diabetes. The burdens of these hospitalizations are distributed unevenly across AA/PI populations. The first step in reducing disparities in preventable hospitalizations is documenting such disparities. Although disparities between many other racial/ethnic groups in preventable hospitalizations have been noted (9), AA/PI subgroups have not been well represented in this research. As preventable hospitalizations have received considerable policy attention (8,30) and hidden AA/PI subgroup disparities are receiving increasing focus (2,12–14), this is a timely and important issue.