Using Indirect Measures to Identify Geographic Hot Spots of Poor Glycemic Control: Cross-sectional Comparisons With an A1C Registry

In the U.S., diabetes prevalence has doubled over the last two decades, and now nearly 1 in 10 Americans has diabetes (1,2). However, these increases have not been evenly distributed, and some geographic regions face a substantially higher burden of diabetes (3,4). For individuals with suboptimal glycemic control, the risk of diabetes-related complications is higher, which leads to frequent hospitalizations, more health care expenditures, disability, and premature death (5,6).

To more effectively reduce the burden of diabetes-related health problems, interventions could target those geographic regions where more people with diabetes have poor glycemic control (7,8). These health interventions for diabetes could also be tailored to the specific health needs of affected communities (9,10). However, the first step is to identify the geographic areas where the proportion of people with poorly controlled diabetes is particularly high.

In New York City (NYC), the A1C Registry was created in 2006 by the NYC Department of Health and Mental Hygiene as a comprehensive collection of all A1C test results (excluding point-of-care tests) from NYC residents (11). Using address information provided with results, the A1C Registry has been used to identify geographic areas with high incidence of poor glycemic control (12,13). Identification of these areas provides potential focal points to target limited resources for improving glycemic control among adults with diabetes. Other efforts using A1C data from diabetes registries have been helpful in exploring disparities in diabetes management. Prior studies have used this data to assess barriers to care and examine the impact of the food and built environment on glycemic control (14–16).

However, most of the country does not have access to a comprehensive registry of A1C values with detailed geographic data. Therefore, the objective of our study was to identify alternative data sources and measures that are more widely available to other jurisdictions. These proxies could be used as indirect measures of identifying hot spots of poor glycemic control. Existing literature has found that glycemic control varies based on race and socioeconomic status (17–19). In addition, a previous study has shown that local geographic regions with higher rates of diabetes-specific emergency department use overlap with areas that have higher rates of microvascular diabetes-related complications (20).

We hypothesized that some of these measures could be used as proxies for poor glycemic control if the geographic hot spots of these factors overlapped substantially with a high prevalence of high A1C values. To test this hypothesis, we compared the accuracy of using hot spots of demographic, socioeconomic, and health care utilization data to identify hot spots of poor glycemic control.

Our study included all unique adults identified in the NYC A1C Registry with likely diabetes (referred to as diabetes going forward) based on having an A1C of ≥6.5% (48 mmol/mol) for two or more tests (12). For health care utilization data, we included data from general, acute-care hospitals in NYC for patients who either had an inpatient hospitalization or emergency department visit and who had a diagnosis of diabetes. We excluded patients from correctional facilities and nursing homes to study the general, noninstitutionalized adult population in NYC. In addition, we excluded data from specialty hospitals (e.g., surgical subspecialty, oncology, or veterans). We only included adults residing in NYC who had a geocodable home address that matched to a NYC census tract (21).

We used laboratory data from the NYC A1C Registry as a gold standard for identifying areas with a high proportion of individuals with poor glycemic control (11). Data on all A1C values (regardless of the indication for the test) are electronically reported from laboratories to the NYC Department of Health and Mental Hygiene. For this study, we used A1C values from 2011 to 2013 to determine whether a person with diabetes had poor control and data from 2006 to 2013 (providing an additional 5-year look-back period) to determine whether a person in our study had diabetes based on a history of two or more A1C tests of ≥6.5% (48 mmol/mol). For sociodemographic predictors, we used the American Community Survey (ACS), which is an ongoing nationwide survey performed by the U.S. Census Bureau. It gathers data including demographic and socioeconomic characteristics. For obtaining census tract–level estimates, survey data are pooled over 5 years (22). To best match the 2011–2013 study period, we used ACS data from 2009 to 2013.

For health care utilization metrics, we used the Statewide Planning and Research Cooperative System (SPARCS) database, which is a comprehensive all-payer New York State claims database (23,24). Using data from 2011 to 2013, we calculated rates of emergency department visits, inpatient hospitalizations, and previously validated estimations of diabetes prevalence based on emergency claims data (25). To compare the study populations of adults with diabetes in the A1C Registry and SPARCS inpatient and emergency patient populations, we abstracted available data on age, sex, race/ethnicity, and health insurance status for analysis in summary statistics.

Our main outcome was the identification of hot spots where the prevalence of poor glycemic control demonstrated statistically significant clustering. To calculate the prevalence of poor glycemic control using A1C Registry data, we divided the number of adults with diabetes whose most recent A1C result was >9.0% (75 mmol/mol) by the total number of adults with diabetes as determined by having two or more A1C results ≥6.5% (48 mmol/mol). An individual was considered to have poor glycemic control if his/her most recent A1C test result in the 2011–2013 study period was >9.0% (75 mmol/mol) (26). This calculation was performed for each census tract in NYC. The hot spot analysis described below was then performed to identify statistically significant clusters of poor glycemic control (20).

We analyzed the selected demographic and socioeconomic factors from the ACS, in addition to health care utilization patterns from SPARCS, to predict the geographic location of poor glycemic control hot spots (18,27). For demographic and socioeconomic characteristics, ACS data were used to calculate the proportion of non-Hispanic black, Hispanic, or minority (nonwhite or Hispanic) residents; the proportion of residents living below 100% of the federal poverty level; and the proportion of residents having less than a high school degree or equivalent. In addition, we used median household income to identify clustering of low median income.

For health care utilization patterns, we identified areas with higher rates of per capita diabetes-specific inpatient hospitalizations and emergency department visits, in addition to estimates of age-adjusted diabetes prevalence based on emergency department claims data. To create the per capita measures, we used SPARCS data for the annual number of inpatient hospitalizations and emergency department visits with a primary diagnosis (not secondary) starting with the prefix of 250 (first three numbers/letters of the code) and divided it by the total number of estimated adults from the ACS data by census tract.

We also used emergency department claims data to estimate the prevalence of diabetes among unique NYC adults who had visited an emergency department at least once during the study period (28). We divided the number of these unique NYC adults in the emergency department claims data who had ever received a primary or secondary diagnosis of diabetes (prefix 250) by the total number of unique adults who had ever visited the emergency department. This estimate was then age adjusted using four age strata (18–24, 25–44, 45–65, and ≥65 years old) and the direct method described by the Centers for Disease Control and Prevention (29). As previously published, this measure demonstrates potential to estimate the prevalence of diabetes at a smaller, more specific geographic level (28).

To identify statistically significant hot spots of poor glycemic control and predictor variables, we used advanced geospatial clustering analysis (20). We first determined the optimal distance band for identifying the maximal spatial autocorrelation or clustering of measured values (30). To do this, we used the Getis-Ord General G statistic within fixed distance bands in 0.5-mile increments from 0.5 to 5.0 miles. Spatial weights were row standardized in this analysis to account for the nonuniform geographic distribution of NYC census tracts (31). After identifying the optimal distance band to perform our hot spot analysis, we used the Getis-Ord Gi* statistic to identify statistically significant hot spots of our main outcome and predictor variables at a 95% confidence level using a false discovery rate correction to account for multiple comparisons. A sensitivity analysis was also performed around the optimal distance band to test the influence of using smaller versus larger geographic distances to identify these hot spots.

To compare the accuracy of our predictor variables, we assessed the degree of overlap at the census tract level between poor glycemic control hot spots as identified with A1C Registry data (as the gold standard) versus hot spots for each of the predictor variables included in the study. Using 2 × 2 tables, we calculated the sensitivity [true positives/(true positives + false negatives)], specificity [true negatives/(true negatives + false positives)], positive predictive value (PPV) [true positives/(true positives + false positives)], negative predictive value (NPV) (true negatives/[false negatives + true negatives]), and overall accuracy [true positives + true negatives)/all observations] for identifying a hot spot of poor glycemic control using the hot spots of predictor variables. CIs for these statistics were determined using the binomial exact method.

To limit the influence of small cell counts, we excluded 57 census tracts that had an adult population estimate <100 for elimination of primarily nonresidential areas (e.g., parks and airports). In addition, we excluded 25 of the remaining 2,110 census tracts that had <30 unique adults in our analyses of A1C Registry data and in our estimates of age-adjusted diabetes prevalence, estimated from emergency department visit data. These exclusions were performed to reduce the influence of a few census tracts that did not have an adequate number of observations to produce reliable estimates (21).

Statistical analyses were performed in Stata 14.2 (StataCorp, College Station, TX). Geographic analysis was performed using ArcGIS Desktop 10.3.1 (Esri, Redlands, CA).

Our study protocol was approved by the institutional review boards at the New York University School of Medicine and NYC Department of Health and Mental Hygiene.

Using the NYC A1C Registry, we identified 570,645 unique adults with diabetes, 101,643 of whom had an A1C >9.0% (75 mmol/mol). There were more adults who were between 25 and 64 years old or male in the group with poorly controlled glucose compared with all adults with diabetes (Table 1). Using emergency department claims data, we also identified 291,515 unique NYC adults with diabetes who had at least one inpatient hospitalization and 393,579 unique NYC adults with diabetes who had at least one emergency department visit. We found that the emergency department population had a slightly higher proportion of non-Hispanic black and Hispanic individuals and younger patients compared with the inpatient hospitalized population (Table 1).

With use of the Getis-Ord General G statistic, maximal spatial clustering for the prevalence of high A1C values was identified within a distance band of 1.5 miles in NYC. For predictor variables, maximal spatial clustering was also identified at 1.5 miles for minority race/ethnicity, poverty, low education, low median income, and age-adjusted diabetes prevalence estimated from emergency department visit data. Maximal spatial clustering for non-Hispanic black residents and Hispanic residents and the frequency of diabetes-specific inpatient hospitalizations and emergency department visits were found at a slightly wider distance band of 2.0 miles. Given these results and slightly higher precision of a smaller distance band, we decided to map all hot spot analyses using 1.5 miles to maintain consistency in our analyses. Using A1C Registry data, we identified hot spots in 42.8% of the 2,085 NYC census tracts analyzed, with the average prevalence of poorly controlled diabetes of 20.6% in hot spots compared with 14.7% in non–hot spots.

Among demographic predictors, we found that the highest sensitivity at 83.0% was achieved by using hot spots of minority residents to identify poor glycemic control hot spots as measured by the gold standard A1C data. Specificity was 88.6% with PPV and NPV at 84.5% and 87.4%, respectively. Comparatively, hot spots of non-Hispanic black residents and hot spots of Hispanic residents were much less sensitive at 58.0% and 51.9%. However, hot spots of non-Hispanic black residents did have a higher specificity, at 91.2%, for identifying hot spots of poor glycemic control (Table 2).

Compared with the proportion of minority residents, socioeconomic characteristics were not as accurate in identifying hot spots of poor glycemic control. The sensitivity of hot spots of low income, poverty, and low education for identifying poor glycemic control hot spots was 65.9%, 59.3%, and 56.4%, respectively. However, hot spots of poverty did demonstrate a high specificity, at 91.2%, for hot spots of poor glycemic control. The PPVs and NPVs for socioeconomic predictors were generally lower than values for the proportion of minority residents (Table 2).

As for health care utilization patterns, the frequency of diabetes-specific inpatient hospitalizations and emergency visits had higher specificity, at 98.8% and 98.9%, for identifying hot spots of poor glycemic control. Sensitivity for these two measures was also higher than any other individual demographic or socioeconomic predictor except for the proportion of minority residents. Hot spots of age-adjusted diabetes prevalence based on emergency department claims data had slightly lower specificity, at 91.7%, than other health care utilization measures. However, hot spots of estimated age-adjusted diabetes prevalence based on emergency department surveillance had the highest sensitivity, at 86.5%, of any of the studied predictors for identifying hot spots of poor glycemic control (Table 2).

Since the frequency of diabetes-specific emergency department visits had the highest PPV at 98.1% and overall accuracy at 89.6%, we mapped the geographic overlap of poor glycemic control hot spots based on A1C Registry data versus hot spots of diabetes-specific emergency department use (Fig. 1). Though the NPV for this measure, at 85.3%, was not the highest of all studied predictors, most of the same areas were identified in a comparison of hot spots in these maps.

We also analyzed the effect of changing the distance band on the sensitivity and PPVs for identifying hot spots of poor glycemic control. Increasing the distance band generally increased sensitivity for most of the predictive factors. Increasing the distance band did not have a consistent effect on PPVs and for most factors was slightly lower with a distance band of 2.5 miles. Decreasing the distance band decreased sensitivity and PPVs, which was most pronounced when we decreased the distance band to 0.5 miles. Increasing and decreasing the distance band also did not have a material effect on which factors had the highest sensitivity or PPVs (Supplementary Tables 1 and 2).

Overall, we found that hot spots of health care utilization measures had higher accuracy than demographic and socioeconomic factors for identifying hot spots with a high prevalence of poor glycemic control. Through this ecologic study, we demonstrate that some of these indirect measures can identify local hot spots of poor glycemic control.

For the greatest impact on health and to address disparities, it is critical to ensure better health intervention allocation to areas that are at greatest risk for the worst health outcomes (32). With geographically detailed data, it is possible to find these areas, which can inform how to tailor interventions that are specific to the neighborhoods with the greatest risk of diabetes-related complications (33).

The accuracy of socioeconomic factors, including median household income, the proportion of residents below the poverty level, and the proportion of residents with less than a high school degree or equivalent, ranged from 70% to 78%. Thus, hot spots of these socioeconomic measures had significantly less overlap with poor glycemic control hot spots than certain demographic factors such as the proportion of minority residents, which had an accuracy of 86%.

However, all health care utilization measures analyzed in our study demonstrated higher accuracy at 88–89% compared with the selected demographic and socioeconomic factors. Of all measures in our study, the frequency of diabetes-specific emergency department visits had the highest PPV, although the frequency of diabetes-specific inpatient hospitalizations essentially mapped the same regions. Also, estimated age-adjusted diabetes prevalence based on emergency department claims data had the highest sensitivity for identifying poor glycemic control hot spots based on A1C data. Though the A1C Registry and SPARCS data likely have divergent biases (individuals of low socioeconomic status are often less likely to have A1C values checked but are more likely to have emergency department visits), there was convergence in the identification of poor glycemic control hot spots between these two data sources.

In the existing literature of population-based studies of A1C values, a national study of 1,350 adults with diabetes from National Health and Nutrition Examination Survey (NHANES) 2007–2010 data demonstrated that poor glycemic control was more common among non-Hispanic black and Hispanic populations (18.7 and 18.8%) versus non-Hispanic white populations (10.1%) (17). The study also found that poor glycemic control was more common among those without a usual source of medical care (22.4%) or among those using hospitals or emergency departments for their health care needs (22.9%) compared with those accessing care at a clinic (15.2%) or a doctor’s office (11.2%) (17). No statistically significant or consistent pattern of association was found between poor glycemic control and education level or poverty income ratios (17).

Our study is not only consistent with prior studies of glycemic control but also provides evidence that in the absence of a population-level A1C Registry with geographically detailed data, health care utilization measures may potentially provide the best proxy for identifying local hot spots of poor glycemic control (18,27). The majority of states across the country already have a systematic process for collecting administrative data on emergency department visits and inpatient hospitalizations (34). Currently, 36 states contribute data to the State Emergency Department Databases (SEDD) and 48 states contribute data to the State Inpatient Databases (SID), both of which are coordinated by the Agency for Healthcare Research and Quality (35).

In areas where comprehensive systems to collect administrative data are not already in place, it may only be necessary to obtain data from selected health care facilities that have the highest geographic coverage of a particular region (25). As shown in a prior study, specific hospitals may be able to act as sentinel hospitals that can provide adequate health surveillance for a particular region as long as the hospital or group of hospitals have a patient population that mirrors the general population for the geographic area of interest (25).

By mapping of hot spots of diabetes-specific health care utilization, these methods have the potential to provide actionable evidence to both public health and health care institutions (36). Population health interventions can then be targeted to places with the highest burden of poor glycemic control (12). With address-level data, it is possible to find these hot spots with a higher level of geographic detail (28). However, our sensitivity analyses suggest that there may be a limit to these methods, as accuracy tends to decrease with use of very small distances.

With the constant pressure of limited resources, it is important to concentrate efforts to improve the health of adults with diabetes in areas with a higher burden of disease, which in part can be characterized by poor glycemic control (37,38). Thus, the PPVs of these indirect measures for tracking poor glycemic control hot spots may be more important if decisions are made to allocate additional resources to areas. This approach would reduce the likelihood of focusing resources in areas (false positives) that may not actually have the anticipated burden of disease. If these methods are used in areas outside of NYC, it should be noted that a higher or lower prevalence of poor glycemic control will have an effect on values for PPV and NPV.

In addition, a modeling approach that combines several predictive factors may be more accurate, though parameters of such models developed in specific cities may not be generalizable to other regions. Also, our geospatial analysis was performed on proportions, rates, and aggregated values. Though some areas may not score high on these metrics, the number of people with poor diabetes control in these areas may still be high on an absolute scale. Therefore, in areas with substantial variation in population density, it may be necessary to perform other types of clustering analyses using count data.

Furthermore, the importance of geographically identifying these poor glycemic control hot spots is important for identifying which particular communities may face a higher burden of diabetes (39). The cultural characteristics of residents in poor glycemic control hot spots could be used to tailor health interventions for diabetes to match the specific cultural context of these identified neighborhoods (40,41). Our finding that nearly half of the NYC census tracts were poor glycemic control hot spots reflects not necessarily that diabetes control is much worse in NYC but, rather, that there is substantial clustering of poor glycemic control in certain neighborhoods.

Further study is needed to understand whether our findings can be generalized to other geographic areas and when these alternative measures may fail to act as good proxies of poor glycemic control. NYC is a dense and unique urban environment with substantial demographic and socioeconomic segregation, in addition to a large undocumented population—all factors that have an impact on access to care, hospital utilization, and the likelihood of having tests like an A1C performed. The next steps in this line of research include conducting validation studies in other regions of the country and looking at the specific areas of poor glycemic control that these indirect measures may miss. These approaches have the potential to identify areas where the burden of poorly controlled diabetes is highest and direct resources to residents at higher risk for adverse diabetes-related outcomes.

Identification of unique individuals in the A1C Registry and administrative data was based on matching algorithms that used a combination of patient variables (e.g., the SPARCS database uses the first two characters of first and last names, last four digits of social security numbers, sex, and date of birth to probabilistically match claims by the same patient). Incorrectly coded or missing data may have led to incorrect matches or double counting of individuals using the algorithms that matched visits or registry values for the same person. If individuals with diabetes or poor glycemic control were more or less likely to be double counted because of these errors, then that would have erroneously inflated or deflated the rates of diabetes or poor control. In addition, hospitals may not always accurately record diagnosis codes, although one study suggested that a diabetes diagnosis code among emergency department visits has a 95% sensitivity and 99% specificity for identifying individuals with diabetes among emergency department patients (42).

Also, we did not have race/ethnicity data on patients within the A1C Registry. Some groups may have different normal ranges for this test or have a higher prevalence of conditions (e.g., sickle cell) which may give false negative results. In NYC, we found that identification of hot spots appeared to work better in some counties compared with others, which may be due to variation in the quality of data provided by different hospitals. NYC is also a dense urban environment, and many of its census tracts contain a high number of residents. Less densely populated regions may require larger geographic units or longer time periods to obtain the counts necessary to provide an accurate analysis. Finally, our study is observational; thus, associations identified cannot be considered as evidence of causation.

Funding. This study was funded by National Institute of Diabetes and Digestive and Kidney Diseases grant K23-DK-110316 to study the geographic distribution of adults with diabetes.

The content of this article solely reflects the opinions of the authors and does not represent the official views of the NYC Department of Health and Mental Hygiene or the New York University School of Medicine.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. D.C.L. researched data and wrote the manuscript. Q.J. researched data and reviewed and edited the manuscript. B.P.T., B.E., C.A.K., and K.J.K. reviewed and edited the manuscript. W.Y.W. wrote the manuscript. D.C.L. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.