Geographic disparities and predictors of vaccination exemptions in Florida: a retrospective study

Study area

This retrospective study was conducted in Florida, which consists of 67 counties, both rural and urban (Fig. 1). Miami-Dade County is the most urban and most populated with approximately 2.7 million residents while Liberty County is the most rural and least populated county with approximately 8,300 residents (US Census Bureau, 2018). Florida population is comprised of 53.2% non-Hispanic White, 26.4% Hispanic, 3% Asian, and 16.9% Black or African American. The remaining 0.5% of the population are Native Hawaiian or Other Pacific Islander or two or more races (US Census Bureau, 2020). The population is 51.1% female and 48.9% male and has the following age distribution: <5 years (5.3%), 6–18 years (19.7%), 19–64 (54.1%), and 65 years and older (20.9%) (US Census Bureau, 2020).

Data sources and cartographic displays

County-level vaccination exemption and population data for 2014 were downloaded from Florida Health CHARTS (Florida Department of Health, 2020). Cartographic boundary files, used for creating maps, were obtained from the United States Census Bureau’s TIGER files (United States Census Bureau, 2019) and the State of Florida Geographic Data Portal (Southwest Florida Water Management District, 2019). Vaccination exemptions per 10,000 population at the county level were calculated and displayed in choropleth maps using ArcGIS (ArcGIS desktop version 10.7; Esri, Redlands, CA, USA). Critical intervals of the choropleth maps were determined using Jenk’s optimization classification scheme.

Predictors of total and non-medical county-level vaccination exemption risks

A conceptual model was used to identify potential county-level predictors of vaccination exemption risks (Fig. 2). Based on the conceptual model, a total of 16 potential predictors were assessed for potential associations with either total and non-medical vaccination exemption risks (Table 1). Predictors of total and non-medical county-level vaccine exemption risks were investigated using two sets of Poisson models for the two outcomes, total and non-medical vaccination exemptions. The first step in building the models involved fitting univariable models between each of the potential predictors and each of the outcome variables (number of total and non-medical vaccination exemptions) with population used as the offset. Univariable associations were assessed at a liberal p-value of 0.2 and variables with a p < 0.2 were considered for further investigation in step two. To minimize potential multicollinearity, two-way Spearman rank correlation coefficients were computed among variables that had p < 0.2. Only one of a pair of highly correlated variables (i.e., with r > 0.7) was retained for assessment in step 2 of the modeling process. The decision on the variable to retain was guided by biological and statistical considerations.

The second step in the modeling process involved fitting multivariable Poisson model using backwards elimination with the predictor variables that had p < 0.2 in step 1 but this time using a critical p-value of 0.05. Confounding was assessed by examining whether the removal of a predictor variable resulted in a >20% change in the coefficients of any other variables in the model (Dohoo, Martin & Stryhn, 2012). Identified confounders were retained in the models regardless of their statistical significance. Overdispersion was assessed by comparing model deviance and degrees of freedom. The Poisson models showed evidence of overdispersion and, therefore, negative binomial regression models were fit to the data using the same process outlined above for the Poisson model. All statistical analyses were performed in STATA version 16.1 (StataCorp, 2019). Overall goodness-of-fit of the final negative binomial models were assessed using Deviance and Pearson Chi-square goodness-of-fit tests. Spatial autocorrelation of the deviance, Pearson and Anscombe residuals of the negative binomial models were assessed using Moran’s I implemented in GeoDa (Anselin, 2019). Statistical significance of the computed Moran’s I values were assessed using permutations approach based on 999 permutations.

Ethical statement

This study was reviewed by the University of Tennessee, Knoxville Institutional Review Board (IRB Number: UTK IRB-20-05957). The board determined that it did not involve human subjects and, therefore, did not require IRB oversight.

Geographic patterns of vaccination exemptions

Total exemptions varied from 0 to 30.16 per 10,000 population (Fig. 3). Nine counties had exemption risks in the top two classes (10.35–15.85 and 15.86–30.16 exemptions per 10,000 population) (Fig. 3). These counties were distributed in five distinct areas of the state: Western Panhandle (Santa Rosa County), central northern area (Columbia, Union, and Gilchrist counties, central (Osceola County)), South-eastern coastal area (Palm Beach, Martin, and St. Lucie counties), and the southern coastal border of the state (Miami-Dade County) (Figs. 1 and 3). Areas with the lowest exemptions (0–3.48 per 10,000 population) were primarily seen in the central portion of the state (Hardee, Desoto, Glades, Okeechobee counties), in the central panhandle (Holmes, Washington, Jackson, Gadsden, Liberty, and Gulf counties), and along north-western coastal and northern parts of the state (Jefferson, Taylor, Lafayette, Dixie, Levy, Suwannee, and Hamilton counties). Two other counties in the northern central part of the state also had low risks of total exemptions (Bradford and Sumter counties).

Non-medical exemptions varied from 0 to 10.40 per 10,000 population (Fig. 3). Fifteen counties had exemption risks in the top two classes (3.71–5.60 and 5.61–10.40 exemptions per 10,000 population) (Fig. 3). These counties were distributed in six distinct areas of the state: Western Panhandle (Santa Rosa and Okaloosa Counties), Central Panhandle (Franklin County), Northeastern (Flagler and Columbia Counties), Central-eastern coastal area (Indian River, Brevard, Martin, and St. Lucie Counties), Central-western coastal area (Pinellas, Hernando, and Citrus Counties), and the South-western coastal border of the state (Monroe, Collier, and Charlotte Counties) (Figs. 1 and 3). Areas with the lowest non-medical exemptions (0–1.20 per 10,000 population) were primarily seen in the central portion of the state (Hardee, Desoto, Hendry, and Okeechobee Counties), in the central panhandle (Holmes, Washington, Gadsden, Liberty, and Gulf counties), and along the northern gulf and eastern panhandle area (Jefferson, Taylor, Dixie, Levy, Lafayette, Suwannee, and Hamilton counties). One other county in the northern central part of the state also had low medical exemptions (Bradford County).

Predictors of total vaccination exemptions

Results of univariable and multivariable Poisson models investigating predictors of total exemptions are shown in Tables 2 and 3. Since the results of the multivariable Poisson model showed evidence of overdispersion (Deviance/df = 29.7; Pearson Chi-square/df = 34.02; Likelihood Ratio Test of overdispersion parameter = 0, p < 0.001), a negative binomial model was more appropriate for these data and hence the results of investigation of predictors of total vaccination exemptions focuses on the negative binomial model results (Table 4). Based on the results of this model, the risk of total vaccination exemptions tended to be higher in counties with more primary care providers per 10,000 population, higher median income, higher percentage of Hispanic population, and higher percentage of population with a college education than in counties with lower values of these predictors (Table 4). There was no evidence of lack of fit of the total exemptions negative binomial model based on the Deviance goodness-of-fit test (p = 0.169). Similarly, there was no evidence of spatial autocorrelation of the deviance residuals (Moran’s I = −0.067; p = 0.272), Pearson residuals (Moran’s I = −0.052; p = 0.353) or Anscombe residuals (Moran’s I = 0.026; p = 0.257).

Predictors of non-medical vaccination exemptions

Findings of the univariable and multivariable Poisson models used to investigate predictors of non-medical exemptions are shown in Tables 5 and 6. As was the case with the total exemptions model, there was evidence of over-dispersion (Deviance/df = 13.98; Pearson Chi-square/df = 16.28; Likelihood Ratio Test of over-dispersion parameter = 0, p < 0.001) of the final Poisson model (Table 6) used to investigate predictors of non-medical exemptions implying that the Poisson model was not appropriate for the investigation. Therefore, subsequent results presented are based on the final multivariable negative binomial model (Table 7). Based on this model, high risks of non-medical exemptions tended to occur in counties with higher percentage of white population. By contrast, lower risks of non-medical exemptions tended to occur in counties with higher percentages of the population: (a) living in rural areas, (b) with a college education, (c) having high school education, and (d) having less than a high school education. As for the total exemptions model, this model indicated no evidence of lack of fit based on the Deviance goodness-of-fit test (p = 0.153) although the Pearson goodness of fit test indicated that the model did not fit the data well (p = 0.003). There was also no evidence of spatial autocorrelation of the deviance residuals (Moran’s I = 0.011; p = 0.343), Pearson residuals (Moran’s I = 0.036; p = 0.205) or Anscombe residuals (Moran’s I = 0.011; p = 0.075).

Geographic disparities in vaccination exemptions

Predictors of vaccination exemptions

Strengths and limitations

This study has shown evidence of county-level geographic disparities in vaccination exemptions and that these disparities are influenced, at least in part, by socioeconomic and demographic factors. However, this study is not without limitations. Exemption data did not include students in home-school programs and therefore this study could not account for the home-schooled population. Additionally, this analysis was exploratory and only included 1 year of exemption data. Future studies should investigate both temporal trends and geographic disparities in vaccination exemptions and should include more years of data. These limitations notwithstanding, the findings of this study are important for guiding and planning public health programs to reduce vaccination exemptions and improve health outcomes for vaccine-preventable diseases.