Geographic disparities and socio-demographic predictors of pertussis risk in Florida

Study area

This retrospective study investigated pertussis risk in the state of 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 is 75.4% White, 25% Hispanic, and 16.1% Black or African American (US Census Bureau, 2018). The state is 51.1% female and 48.9% male and has the following age distribution: <5 years (5.4%), 5–9 years (5.5%), 10–14 years (5.7%), 15–19 years (5.8%), 20–24 years (6.2%), 25–34 years (12.9%), 35–44 years (12.1%), 45–54 years (13.3%), 55–59 years (6.8%), 60–64 years (6.4%), 65–74 years (10.9%), 75–84 years (6.2%), and ≥85 years and older (2.6%) (US Census Bureau, 2018).

Data sources and variable selection

Yearly number of cases of pertussis for each county for the time period 2010–2018 were obtained from the Florida Department of Health. Cases were aggregated into three time periods: 2010–2012, 2013–2015, and 2016–2018. The 3-year aggregation was necessary due to the very small numbers of yearly cases observed and the fact that pertussis occurrence has a 3–5 year periodicity (Broutin et al., 2010; Clark, 2014; Bouchez & Guiso, 2015; CDC, 2015). Population data were downloaded from Florida Health CHARTS, a community health assessment tool created by the Florida Department of Health (Florida Department of Health, 2019). Vaccination exemption data were also obtained from the Florida Department of Health. All cartographic boundary files used for creating maps were downloaded from the United States Census Bureau’s TIGER files and the State of Florida Geographic Data Portal (Southwest Florida Water Management District, 2019).

County level age-adjusted pertussis risks were computed and presented as number of cases per 100,000 population. A total of 20 potential predictors of pertussis risk were identified and considered for investigation at the county level: education, sex, race/ethnicity, income level, employment, health insurance, number of primary care physicians, percentage of population living in rural areas, vaccination coverage, and vaccination exemptions (Table 1).

Investigation of high-risk clusters of pertussis

Tango’s flexible spatial scan statistic (FSSS) was used to identify the geographic locations of high-risk clusters of pertussis using FleXScan (Tango & Takahashi, 2005). Since the flexible spatial scanning window detects both circular and irregularly shaped clusters (Jacquez, 2009; Odoi et al., 2019; Lord, Roberson & Odoi, 2021), it is ideal for situations like this one involving investigation of clusters whose shapes are unknown at the outset (Roberson et al., 2019). It has been shown that Tango’s flexible scan statistics has good power as well as the ability to detect irregularly shaped clusters more accurately than Kulldorff’s spatial scan statistics (Tango & Takahashi, 2005). Poisson probability model was used specifying maximum spatial scanning window size of 15 counties. Clusters were identified based on likelihood ratio tests. For statistical inference, 999 Monte Carlo replications were used, and the null hypothesis of total spatial randomness was rejected if the p-value was ≤0.05.

Investigation of predictors of county-level pertussis risk

Predictors of county-level pertussis risk were investigated using an ordinary least squares regression model. The first step in building the model involved fitting univariable models between each of the potential predictors and the outcome variable (log-transformed age-adjusted pertussis risk). Univariable associations were assessed at a liberal alpha of 0.2 and therefore variables with a p < 0.2 were considered for further investigation in step two. Two-way correlations were run for variables considered for step 2. Only one of a pair of highly correlated variables (absolute value of r > 0.7) were assessed in step two. The decision regarding which of a pair of highly correlated variables to include in step two was based on biological and statistical considerations. The second step involved fitting multivariable ordinary least squares model using manual backwards elimination with the predictor variables identified for inclusion in step one. During this step statistical significance was assessed using an alpha of 0.05. Confounding was investigated by assessing whether the removal of a variable resulted in a >20% change in the coefficients of any other variables in the model (Dohoo, Martin & Stryhn, 2012). If a confounder was identified, it was forced into the model regardless of its statistical relationship with the outcome variable. Robust Lagrange Multiplier (RLM) tests, implemented in GeoDa, were used to assess for spatial autocorrelation in the residuals using queen spatial weight (Anselin, Syabri & Kho, 2016). Multicollinearity was assessed using both Variance Inflation Factor (VIF) in SPSS (IBM SPSS Statistics for Windows, Version 25.0; IBM, Armonk, NY, USA) and the Multicollinearity Condition Number in GeoDa (Anselin, Syabri & Kho, 2006).

Cartographic displays

All geographic information system (GIS) manipulations and cartographic displays were performed in ArcGIS (ArcGIS Desktop 10.6.1. Esri, Redlands, CA, USA). Statistically significant county-level predictors of pertussis risks, age-adjusted pertussis risks, and significant clusters of pertussis risk were displayed in maps. The choropleth maps of pertussis risk and of high-risk clusters were generated for the full study period and three-year time periods: 2010–2012, 2013–2015, and 2016–2018. Critical intervals in maps of the determinants of pertussis risk as well as maps of pertussis risk were determined using Jenk’s optimization classification scheme (Jenks, 1967) for the full study period (2010–2018) and were applied to pertussis risk maps of other time periods for consistency.

Ethical statement

This study was assessed by the University of Tennessee Institutional Review Board (IRB Number: UTK IRB-20-05957) which determined that it did not involve human subjects as defined in 45 CFR 46.102 (e) (1), since it did not involve use of identifiable private information. Thus, the IRB determined that neither IRB review, nor certification of exemption from review, was required.

Geographic and temporal distribution of pertussis risk

County-level pertussis risk varied from 0 to 116.31 cases per 100,000 people during the study period (2010–2018) (Fig. 2A). Four regions with high risks of pertussis were evident: the Pensacola metropolitan area in the western panhandle (Escambia and Santa Rosa counties), the Tampa Bay area (Hillsborough, Pasco, and Polk counties), Jacksonville metropolitan area (Nassau, Duval, and St. Johns counties), and the Cape Coral-Fort Meyers-Naples Area (Collier, Hendry, and Lee counties) (Figs. 1 and 2A). High risks were also observed in Columbia and Suwannee counties. Areas of low risk, on the other hand, were evident around Glades, Union, Liberty, and Calhoun Counties (Figs. 1 and 2A).

The pattern of geographic distribution of pertussis risk was slightly different for the time period 2010–2012, compared to the whole time period (2010–2018), with high risks being observed in Santa Rosa county in the western Panhandle, around the northern gulf coast in Taylor, Lafayette, and Levy counties, in the northwestern region around Jacksonville in Nassau, Baker, Duval, and Putnam counties, and around the Tampa Bay area in Pasco, Hillsborough, and Polk counties (Figs. 1 and 2B). Though the general spatial patterns of pertussis risk for the time period 2013–2015 were similar to those of 2010–2012, a number of counties (such as Columbia, Suwannee, Washington, etc.) had higher overall risks in 2013–2015 than 2010–2012 (Figs. 1 and 2B–2C). In addition, high risks were identified in Gilchrist County, around the Tampa Bay area in Pasco, Hillsborough, Polk, Highlands, DeSoto, Manatee, and Sarasota counties, and around the Cape Coral area in Lee and Collier counties. Although county level pertussis risks generally increased from 2010–2012 to 2013–2015, risk tended to be lower during the 2016–2018 time period compared to the 2013–2015 period (Figs. 2C and 2D). Highest risks were observed in Nassau County followed by Nassau, Hillsborough and Hendry counties.

Clusters of high pertussis risk

A total of 11 clusters were identified for the analysis that included data for the whole study period (2010–2018) (Table 2 and Fig. 3A). The primary cluster during this time period included one county (Hillsborough) and had a risk ratio (RR) of 2.1. The risk of pertussis in this cluster was 43.1 cases per 100,000 and it had a total of 556 observed cases of pertussis when only 264 cases were expected. Based on geographic size, the largest cluster was Cluster 3 which included two counties (Polk and Pasco). The risk of pertussis in this cluster was 1.51 times higher than the risk outside the cluster. A total of 340 pertussis cases were observed in this cluster when only 223 cases were expected. It is worth mentioning that the cluster with the highest RR (Cluster 2) included only one county (Columbia) and had a RR = 5.7, a risk of 116.3 cases per 100,000 persons, 79 observed cases while only 13 were expected. Counties in the Tampa Bay area (Hillsborough, Polk, and Pasco counties) were consistently included in high-risk clusters throughout the study period (Fig. 3A).

Seven clusters, having RRs ranging from 1.6 (in cluster 7) to 3.4 (in cluster 4), were identified during the time period 2010–2012 (Table 3 and Fig. 3B). The time period 2013–2015, on the other hand, had a total of 11 clusters with RRs ranging from 1.4 (in cluster 9) to 12.4 (in the primary cluster) (Table 4 and Fig. 3C). The latter was located in Columbia County and had the highest RR across all time periods. Collier county, in the Naples metropolitan area, and Nassau county, in the greater Jacksonville area, were also identified as parts of high-risk clusters during the time periods 2013–2015 (Fig. 3C) and 2016–2018 (Fig. 3D). The fewest number of clusters were identified in the 2016-2018 time period (Table 5 and Fig. 3D). This was also the only time period during which a multi-county high pertussis risk cluster (RR = 1.95; p = 0.002) was identified.

Predictors of pertussis risk

Results of descriptive analysis of the potential predictors and assessment of their associations with pertussis risk are presented in Table 6. The following variables were assessed in the multivariable model: rural status, education variables, percent female, primary care providers per 100,000 population, median income, percent vaccinated, percent with medical exemption, and percent with religious exemptions. Only three of these variables were significant in the final model: rurality, county level median income and percentage of females in the county. County level pertussis risks tended to be high in rural counties (p = 0.021), those with higher median income (p = 0.039), and those with higher percentages of females (p < 0.001). It is worth pointing out that % female was a distorter variable for the association between rurality and risk of pertussis with the coefficient of the rural variable switching from a negative association with log-risk of pertussis, in the univariable model, to a positive association in the final model when percent females is added to the model (Tables 6 and 7). There was no evidence of multicollinearity as all the VIFs were <4 and the multicollinearity condition number was <20 (Table 7). Additionally, there was no evidence of non-normality of the residuals (Jarque-Bera test p = 0.678). Both the Robust Lagrange Multiplier tests for lag (LM_lag p = 0.585) and error (LM_error p = 0.472) showed no evidence of spatial dependence of the residuals.

Geographic distribution of significant predictors of pertussis risk

Geographic distributions of the three significant predictors of county-level pertussis risk are shown in Figs. 1 and 4. Unsurprisingly, the general distribution of these predictors is similar to the geographic distribution of pertussis risk in Florida between 2010 and 2018 (Fig. 2A). Santa Rosa and Okaloosa counties in the western panhandle, Nassau, St. John’s, and Clay counties in the Northwestern coastal area, Hillsborough, Manatee, and Sarasota counties in the Tampa Bay area, and Collier and Monroe counties in the Fort Myers area all had relatively high median incomes (Figs. 1 and 4A). These counties were also consistently identified as counties with high pertussis risk and in high risk clusters (Figs. 2A, 3A and 4A).

Counties with higher percentages of females are evident in the central panhandle (Gadsden and Hernando counties) and along the eastern and western coastal borders. Counties with greater than 52% female population included: Gadsden, Leon, Sarasota, Hernando, and Pinellas counties (Figs. 1 and 4B). All of these counties had high pertussis risk (>8 per 100,000) and many counties with higher female populations were identified in high risk clusters, including Hillsborough, Lee, Escambia, Santa Rosa, Duval, St. John’s, Polk, and Pasco counties. Finally, rural counties, were predominately located in the central and eastern panhandle and the central body of the state (Fig. 1).

Predictors of pertussis risk

Significant predictors of county-level pertussis risk identified in the current study were percentage of females, median income, and rurality. Counties with high proportions of females tended to have high pertussis risk. This is consistent with findings from individual level studies that reported higher risks of pertussis among females than males in the United States, Canada, and Poland (De Serres et al., 2000; Paradowska-Stankiewicz & Rudowska, 2015) . An analysis of an outbreak of pertussis in Quebec, Canada, reported that the main risk factors for pertussis were being female and working in an educational or healthcare setting. The authors proposed that the higher risk may be attributed to greater awareness of pertussis among females which could lead them to seek medical care more readily than their male counterparts (De Serres et al., 2000). Higher pertussis risk and pertussis complications among females have also been reported in the United Kingdom and Canada (Jenkinson, 1995; Abu-Raya et al., 2020). Since the current study is not an individual-level study, the findings of the above individual-level studies cannot be directly implied to it. However, it is possible that the associations seen at the individual-level may apply at the county-level as well. Suffice it to say that this is the first ecological study that has investigated sex as a potential sociodemographic predictor of pertussis risk despite evidence at the individual level that sex is an important predictor of pertussis risk (Broutin et al., 2004; Huang et al., 2017; Alimohamadi et al., 2020).

The association between higher median income and higher county-level pertussis risk may be related to vaccine refusal since higher median income has been well-documented as associated with vaccine refusal for ideological reasons (Wei et al., 2009; Remes et al., 2014). However, it is worth noting that vaccination delay has also been identified as a significant contributor to disparities in vaccination coverage and often occurs along socioeconomic and demographic lines. Children from economically disadvantaged families, with poor access to healthcare services, often have delays in their vaccination schedules. This is of particular interest because this delay can increase risk of not only pertussis illness but other vaccine-preventable diseases as well (Luman et al., 2005; Feemster et al., 2009; Schaller, Schulkind & Shapiro, 2019). However, children with vaccine delays are usually fully vaccinated by the time they begin attending school since it is required unless an exemption has been approved (Schaller, Schulkind & Shapiro, 2019). This may explain why the current study did not identify an association between vaccination coverage and county-level pertussis risk, since we could not assess the vaccination coverage of children prior to starting compulsory kindergarten education.

While high-risk clusters were identified around metropolitan areas, the final multivariable model indicated that risk of pertussis was higher in rural counties than urban counties. These contradictory findings may be the result of distortion effects of percent females in the model. However, our finding that rurality is a predictor of higher county-level pertussis risk is consistent with reports of a geospatial analysis of pertussis in Saskatchewan, Canada, that rural areas had higher pertussis risk than non-rural areas (Medu et al., 2018). There is a clear need to explore the relationship between population density, rurality, and risk of pertussis more deeply, as there is conflict not only in this study but also in the literature. Possible explanations for these findings include lower vaccination rates due to lack of access to healthcare services or geographical differences related to objections to vaccinations.

Strengths and limitations

To our knowledge, this is the first study to investigate spatial patterns and clusters of pertussis in Florida using rigorous spatial epidemiologic methods. The FSSS has several strengths, over other methods such as the Kulldorff’s spatial scan statistics and local indicators of spatial associations (LISA) methods such as Moran’s I: (a) it does not have the problem of multiple comparisons associated with LISA; (b) does not have pre-selection bias; (c) it is able to detect irregularly shaped clusters since it identifies all cluster shapes as opposed to Kulldorff’s spatial scan statistics which only identifies circular or elliptical clusters (Tango & Takahashi, 2005). Moreover, the Tango’s FSSS has been shown to work better than Kulldorff’s scan statistic for identifying small clusters (Tango & Takahashi, 2005). County level spatial analysis was necessary to avoid the small number problem associated with analysis at lower geographical scales because of the small numbers of cases of pertussis. The identified spatial clusters and predictors are useful for guiding health planning and policy. In addition to identifying county-level predictors of pertussis risk, this study has demonstrated the usefulness of using spatial scan statistic to identify areas where preventive resources are needed through the identification of high-risk clusters of pertussis. However, this study is not without limitations. Pertussis is an under-reported disease and, therefore, the county-level pertussis risk reported in this study is likely lower than the true risk. Due to the small number of cases and to protect anonymity and confidentiality of cases, only yearly pertussis data were used. These data were further aggregated into 3-year time periods for descriptive investigation to ensure stable estimates due to the small yearly case counts. Future studies will need to investigate individual level data covering longer study periods. These limitations notwithstanding, the findings of this study are important for guiding and planning public health programs.