Predictors of county-level diabetes-related mortality risks in Florida, USA: a retrospective ecological study

Ethics approval

This study was approved by the University of Tennessee Institutional Review Board (IRB number: UTK IRB-23-07809-XM). The review board determined that the study “is eligible for exempt review under 45 CFR 46.101 pursuant to category 4ii: Secondary research for which consent is not required: Secondary research uses of identifiable private information or identifiable biospecimens, if information, which may include information about biospecimens, is recorded by the investigator in such a manner that the identity of the human subjects cannot readily be ascertained directly or through identifiers linked to the subjects, the investigator does not contact the subjects, and the investigator will not re-identify subjects”. The authors did not have access to information that could identify participants during or after data collection.

Study area

This study was conducted in the state of Florida which is located between 27°66′N and 81°52′W and spans 65,758 square miles. County land area ranges from 243.6 square miles (Union County) to 1,998 square miles (Collier County) (United States Census Bureau, 2023a). As of 2022, Florida was one of the most populous states in the US. About 22.2 million people live in the state, 50.8% of whom are female and 49.2% are male (United States Census Bureau, 2023b). Miami-Dade County is the most populous (2.6 million people), whereas Liberty County, located in the northern part of the state (Fig. 1), is the least populated with only 7,987 people (United States Census Bureau, 2023b). The age distribution of the adult population in Florida is 24% 18–34 years old, 26% 35–49 years old, 25% 50–64 years old, and 22% are seniors (≥65 years old) (United States Census Bureau, 2023b). There are 76.8% White, 17% Black, and 6.2% all other races in Florida. Of the 67 counties in the state, 30 are rural and 37 are urban (Fig. 1).

Diabetes-related death data

Individual-level death data covering the time period January 1 to December 31, 2019, were obtained from the Florida Department of Health. The International Classification of Disease (ICD) 10th revision was used to identify the cause of death, and ICD-10 codes E10-E14 were used to identify diabetes-related deaths (World Health Organization, 2019). No distinction was made between type 1 and 2 diabetes. The county of residence of the deceased was used to aggregate the number of diabetes-related deaths to county-level using R statistical software version 4.2.2 (R Core Team, 2023). Population estimates for 2019 were obtained from the American Community Survey (ACS) (United States Census Bureau, 2022) and used as denominator to calculate DRMR. County-level DRMRs were then calculated and expressed as number of deaths per 100,000 persons. A conceptual model of the potential predictors of DRMR was constructed (Fig. 2). Data on potential predictors of DRMRs, for the year 2019, were extracted from several data sources (Table 1).

Descriptive statistics

Descriptive analyses were conducted using R statistical software version 4.2.2 (R Core Team, 2023) and implemented in R studio version 1.4.1717 (R Studio Team, 2020). Normal distribution of continuous variables was evaluated using the Shapiro–Wilk test. Since some of the continuous variables were not normally distributed, medians and interquartile ranges were used for summary statistics (Table 2).

Investigation of county-level predictors of DRMR

A global ordinary least squares regression (OLS) model was used to identify county-level predictors of DRMR in Florida. After selecting potential predictors using the conceptual model, a two-step process was used to fit a multivariable model with the outcome specified as DRMR. The first step in model-building was to assess the univariable associations between each potential predictor and the outcome. A liberal p-value of ≤0.20 was used for this assessment. Using p-value ≤0.20 allows for assessment of potentially important confounders during the multivariable analysis stage (Dohoo, Martin & Stryhn, 2012). All the variables that showed significant associations (based on a relaxed p-value of ≤0.20) in the univariable analyses were subjected to two-way Spearman rank correlation analyses. Spearman rank correlation analysis was appropriate for this assessment because some of the continuous variables were not normally distributed and Spearman correlation analysis does not assume normal distribution (Zar, 2014). When two variables showed a strong correlation (r > 0.7), only one of them was included in the subsequent multivariable model. The decision on which variable to include in the multivariable model from a pair of highly correlated variables was based on biological and statistical considerations. Backward elimination process was then performed, using a critical p-value of 0.05, to fit the final multivariable model. The backward elimination approach allows for assessment of confounders during the modeling process (Dohoo, Martin & Stryhn, 2012), enhances the accuracy of prediction, and reduces the likelihood of overfitting the data (Draper & Smith, 1998). Confounding variables were then evaluated by comparing changes in regression coefficients after running the model with and without the suspected confounder. If there was ≥20% change in the coefficients of any of the variables in the model, the suspected variable was then identified as a confounder and retained in the model regardless of its statistical significance. Two-way interaction terms of the variables of final multivariable model were evaluated based on biological relevance, and only those with a p-value ≤0.05 were included in the final model. The variance inflation factor (VIF) was used to assess multicollinearity in the final model. If the VIF value of a variable was ≥10, it was considered to have high collinearity with at least one of the other variables in the model. Overall goodness-of-fit of the final model was assessed using adjusted R-squared (R²) and Akaike Information Criterion (AIC). Residual plots, Jarque–Bera test, and Breusch-Pagan test were used to evaluate the assumptions of normality and homoscedasticity. The residuals were also used to identify outliers, while leverage, Cook’s distance, and Difference in Fits (DFITS) were used to identify influential observations. All the analyses were performed using R statistical software version 4.2.2 (R Core Team, 2023). Moran’s I, implemented in GeoDa (Anselin, Syabri & Kho, 2006), using 1st order queen contiguity spatial weights, was used to assess for spatial effects or spatial autocorrelation in the residuals.

Cartographic display

Cartographic boundary files were downloaded from the TIGER geodatabase (United States Census Bureau, 2023a) and used to generate maps. QGIS 3.34 (QGIS org, 2024) was used for all cartographic displays. Choropleth maps for the distribution of county-level diabetes-related mortality risks and significant predictors from the final multivariable model were generated using Jenk’s optimization classification scheme.

Geographic disparities in distribution of diabetes-related mortality risks

There was a total of 6,078 diabetes-related deaths reported in Florida in 2019. Of these, 59.3% were male, 40.7% were female, 75.7% were White, 20.4% were Black, and 3.9% were all other races. The percentage of diabetes-related deaths was highest (69.6%) among seniors (≥65 years old). The county-level mortality risks varied across the state, ranging from 9.6 per 100,000 persons in St. Johns County to 75.6 in Desoto County (Fig. 3). Of the 67 counties in the state, 35 had mortality risks equal to or greater than the national average (31.1 per 100,000 persons). Counties with high mortality risks were mainly in the northern (Holmes, Washington, Jefferson, Taylor, Santa Rosa, and Dixie), central (Citrus, Hernando, Sumter, Marion, and Putnam), and south-central (Desoto and Hardee) parts of Florida (Figs. 1 and 3). Conversely, the southernmost counties (Broward, Collier, and Monroe) had relatively lower mortality risks compared to other parts of the state. It is worth noting that most counties with high mortality risks were mainly in rural areas.

Predictors of diabetes-related mortality risks

Results of the univariable associations between county-level DRMR and its potential predictors are shown in Table 3. Forty-six of the 59 variables assessed had significant univariable associations with DRMR using a relaxed critical p-value of ≤0.20. However, based on the final multivariable model, significant predictors of disparities in county-level DRMRs were percentage of population that had diabetes (p < 0.001), were aged 65 or above (p < 0.001), were current smokers (p = 0.032), were insufficiently physically active (p = 0.036), and percentage of households without vehicles (p = 0.022) (Table 4). No significant interactions and confounding were detected. The p-value of Moran’s I (p = 0.747) indicated that there was no spatial effects or spatial autocorrelation in the residuals and hence the model did not violate the independence assumption. A comparison of the observed DRMR and the model estimates of the DRMR is shown in Fig. 4. The general pattern in distribution of the model estimates is quite similar to that of the observed DRMR indicating that the model is doing well in predicting DRMR in the study area (Fig. 4).

The northern, central, and south-central counties had high percentages of population with diabetes (Fig. 5) and insufficiently physically active population, mirroring the spatial patterns of DRMRs (Fig. 3). High percentages of population with diabetes and insufficient physical activity tended to mainly occur in rural counties. The percentage of smokers tended to be high in the northern regions of the state, whereas high percentages of households without vehicles were evident across both northern and southern parts of the state (Fig. 5). Despite lower DRMR in the southernmost counties of the state, these regions had higher percentages of population aged 65 or above compared to other parts of the state (Fig. 5).

Strengths and limitations

This is the first study to investigate county-level predictors of diabetes-related mortality risks in Florida using rigorous statistical approaches. The findings of this study are useful for guiding evidence-based interventions by identifying DRMR disparities and its predictors. This study investigated county-level predictors of DRMR in Florida for one year (2019). It is possible that these associations change over time based on the changes in population socioeconomic and demographic profiles. Therefore, it is important that these investigations be part of regular health surveillance programs to provide the most current information to guide health planning and service provision. However, this study has some limitations. The BRFSS survey data are self-reported, and so might have reporting bias. The survey does not distinguish between type 1 and type 2 diabetes. Although type 1 and type 2 diabetes have different etiologies (type 2 diabetes are more influenced by lifestyle, socioeconomic, and behavioral factors, while type 1 diabetes is due to autoimmune disease), we only considered socioeconomic and demographic factors in this study. However, since the majority (90–95%) of diabetes in the United States is type 2, we expect the fact that there was minimal effect on the results of this study. Finally, since this is an ecological study conducted at the county-level, inferences must only be made at the county level and not individual level to avoid ecological fallacy. These limitations notwithstanding, the findings provide useful information regarding disparities and determinants of DRMR in Florida.