Childhood Brain Cancer in Florida: A Bayesian Clustering Approach

Abstract

In this article, we focus on geocoded data for pediatric brain cancer in Florida. Specifically, we examine zip code level pediatric brain cancer counts from the Florida Association of Pediatric Tumor Programs (FAPTP) childhood cancer registry from 2000-2010 and assess the degree of spatial clustering in these data. We assume Bayesian models for relative risk and examine a variety of posterior measures that indicate excess risk (exceedence probability of relative risk or positive residual). We assume a standard Poisson convolution model and examine a zero-inflated (ZI) model with a factored intercept. We conclude that there is evidence of excess risk in a number of relatively dispersed zip codes across the state but there appears to be some concentration of high excess risk in Polk, Lake, and Sumter counties (west of Orlando and north east of Tampa). These excesses are confirmed across the models.

Discussion: Statistical Cluster Detection, Epidemiologic Interpretation, and Public Health Policy

“…ZCTAs are constructed from complete census regions (e.g., blocks) combined together so that a large majority of residents of a ZCTA share the same ZIP code, but ZIP codes can (and often do) follow different boundaries than Census blocks (see articles by Grubesic andMatisziw 2006 andBeyer, Schultz, andRuston 2008 for more detailed discussion). In the present analyses, it is important to note that the ZCTA data does not fully cover the state of Florida-a particular gap (due to low population density in the Everglades) is seen in the southernmost tip in the maps in figures 1 and 2 in Heaton (2014), figures 1 and 2 in Wang and Rodriguez (2014), figures 1-11 in Lawson and Rotejanaprasert (2014), and figures 1 and 3 in Zhang, Lim, and Maiti (2014). It is important to appreciate that ZCTAs provide a close approximation to link registry-based disease cases to census demographics defining the population at risk, but that it is still an approximation and may merit closer investigation to fully understand potential clusters, particularly those based on small local numbers of individuals at risk.…”

“…Can we define the distribution of values we would expect to see in each location in the absence of clustering, and accurately and reliably detect deviations from this distribution? The approaches used by the authors build on this general motivating question using ideas from hypothesis testing (Amin et al 2014), model-based (posterior) estimation (Heaton 2014;Lawson and Rotejanaprasert 2014;Zhang, Lim, and Maiti 2014), and machine learning (Wang and Rodriguez 2014). Again, all authors are interested in the statistical identification and evaluation of local aggregations of cases in space and time to see if there are any aggregations (clusters) that are inconsistent with known factors driving incidence.…”

“…Some ZCTAs contain larger numbers of individuals at risk, allowing better precision in estimation, but some likely have less than one case expected. All statistical approaches applied in the five articles seek to combine or borrow information between ZCTAs to improve local estimates either through combining data across neighboring ZCTAs (the set of potential clusters evaluated in Amin et al 2014 and the clusters of equal risk in Zhang, Lim, and Maiti 2014), or through spatial smoothing/correlations between ZCTAs (Heaton 2014;Lawson and Rotejanaprasert 2014;Wang and Rodriguez 2014). All authors note that their at-risk population values are from the Census and it appears they limit these to age groups consistent with the pediatric at-risk population, but to aid in epidemiologic interpretation of the statistical results, it is important to clarify this detail.…”

Modeling pediatric tumor risks in Florida with conditional autoregressive structures and identifying hot-spots

“…Also, there are several small-sized high risk regions in Northern Florida. Lawson and Rotejanaprasert (2014) applied the standard Poisson convolution (SPC) model to pediatric brain cancer data and considered zero-inflation with a factored intercept. They mapped exceedance probability of relative risk greater than 1, 0 or 2, or residuals greater than 0 for each region with their SPC model and the model by Besag, York, and Mollié (1991).…”

“…Then, we investigated a zero-inflated Poisson normal model to see if we need a zero-inflated component since there are relatively many areas with zero counts. Lawson and Rotejanaprasert (2014) also considered a model that can accom-modate excess zero count in data but their model is different from a zero-inflated model in that zero count information in their model is incorporated in the Poisson mean while a zero-inflated model consider a mixture of a Poisson model and zero mass.…”