Aarya Satardekar

Geographic variation (GV) in screening patterns for breast cancer exists among Zip Code Tabulation Areas (ZCTAs) in Florida, but most spatial analyses are based on frequentist point estimates which do not formally represent uncertainty. This study used a three-stage analytical approach to the breast cancer screening data at the zip code tract (ZCTA) level in Hillsborough County, Florida (n = 55 ZCTAs): first, a frequentist Poisson regression was applied with diagnostics for multicollinearity; second, global spatial autocorrelation (GSA) analysis was conducted using Moran’s I; and third, a Bayesian Poisson and a Bayesian negative binomial regression were performed, both estimated via the No-U-Turn Sampler in the brms/Stan framework. In ArcGIS Pro 3.6, spatial analyses were carried out. The dependent variable was the number of breast cancer screening exams conducted at the ZCTA level over the study period. Across all racial/ethnic subgroups, the observed number of screenings was correlated with the number of females in the household, while no independent correlation was found for median household income, insurance status or age by stratum variables after adjustment. There was no significant and strong spatial autocorrelation across the study area (Moran’s I = 0.003, z = 0.326, p = 0.745). The Poisson model did best among the Bayesian models with a Bayesian R² of 0.91, RMSE of 5.40, and MBE of 0.02. The results show the usefulness of Bayesian uncertainty quantification in small area public health surveillance and offer a framework for quantifying geographic variation in screening activity in a probabilistic manner. The results only compare screening examination (not population-standardized screening rates) and should be considered to reflect screening volume rather than screening participation.

Breast cancer screening patterns exhibit geographic variation across Zip Code Tabulation Areas (ZCTAs) in Florida, yet most spatial analyses rely on frequentist point estimation without formally characterizing uncertainty. This study applied a three-stage analytical framework to ZCTA-level breast cancer screening data in Hillsborough County, Florida (n = 55 ZCTAs): frequentist Poisson regression with stepwise multicollinearity diagnostics, global spatial autocorrelation analysis using Moran’s I with inverse-distance weighting, and Bayesian Poisson and Bayesian negative binomial regression with Jeffreys non-informative priors estimated via the No-U-Turn Sampler in R (brms/Stan). Spatial analysis was conducted in ArcGIS Pro. Racial and ethnic female population counts for White, Black or African American, and Hispanic or Latino groups were the strongest and most consistent predictors of screening counts. Median household income, insurance status, and age-stratified variables showed no independent association at the ZCTA level. Global Moran’s I was near zero and non-significant (I = 0.003, z = 0.326, p = 0.745). The Bayesian Poisson model showed superior fit compared with the Bayesian negative binomial model (Bayesian R² = 0.91, DIC = 367.2, RMSE = 5.40, MBE = 0.02). These findings associate screening concentration with the geographic distribution of demographic groups and demonstrate the value of a Bayesian uncertainty-oriented framework for small-area public health analysis.

Leukemia is a cancer of the blood and bone marrow, that hinders the normal production of healthy blood cells. In exploring mathematical hypotheses for leukemia, three distinct approaches are proposed. First is an over-dispersed Poisson leukemia regression model, with the consideration of outliers being addressed by applying a negative binomial model featuring a non-homogenously distributed mean. Secondly, an eigenfunction, eigendecomposition spatial filter algorithm is introduced, aiming to identify potential leukemia clusters based on hyper/hypo-endemic aggregation/non-aggregation orientations. Lastly, a Bayesian hierarchical model is advocated for determining causation covariates within a non-frequentistic model. This research examined the spatial aggregation of leukemia cases by utilizing sociodemographic data at the zip code level in Hillsborough County, Florida. The investigation involved spatial autocorrelation and Bayesian analyses to pinpoint the covariates linked to the risk of leukemia. The Poissonian regression model revealed a nondispersed paradigm. Hence, there was no need to utilize the negative binomial regression to treat the outliers. A second-order eigenfunction eigendecomposition revealed multiple non-zero autocorrelated clusters throughout various zip codes in Hillsborough County. The hot spots were 33647, 33578, and 33511 and the cold spots were 33621, 33503, and 33530. Our proposed approach identifies leukemia hotspots among whites and Asians aged 65+. Urban residential communities in 33647 were most vulnerable to leukemia. By leveraging social media and fostering collaboration between public health and technology sectors, we can work towards achievable leukemia prevention for at-risk populations. Future research should explore the method's applicability at the state level and promote routine blood tests among people.