Hierarchical modeling and other spatial analyses in prostate cancer incidence data.

Background: State central cancer registries are often asked to respond to questions about the spatial distribution of cancer cases. Spatial analysis methods and technology are evolving rapidly, and can be a considerable challenge to registries that do not have staff with training in this area. The purpose of this article is to describe a general methodological approach that potentially might be a starting point for many cancer registry spatial analyses at the county level.

Methods: Prostate cancer incident cases (N=31,159) from the Louisiana Tumor Registry from 1988 to 1999 were used for illustrative purposes. To explore spatio-temporal patterns, analyses focused on four time periods, each 3 years in length: 1998-1990, 1991-1993, 1994-1996, and 1997-1999. For each time period, race-specific (white and black), direct age-adjusted incidence rates and indirect standardized incidence ratios (SIRs) were calculated, smoothed using Bayesian methods, and assessed for evidence of spatial autocorrelation using global and local Moran's I. Hierarchical generalized linear models (HGLM) were fitted to identify significant covariates. Clusters of elevated and lower rates were identified using a spatial scan statistic (SaTScan).

Results: Temporal trends in SIRs in both race groups were consistent with the introduction of prostate specific antigen (PSA) testing in Louisiana during the late 1980s and early 1990s, but possibly with a lag in black males. Clusters of lower than expected values were observed for white males in the central (p=0.001) and southeastern coastal areas (p=0.001), and to a greater extent for black males in the central (p=0.001), southwestern and southeastern coastal parishes (p=0.001).

Conclusions: Mapping disease occurrence by time period is an effective way to explore spatio-temporal patterns. HGLM models and software are available to control for covariates and for unstructured and spatially structured variability that may confound spatial variability patterns.