Methodologic implications of social inequalities for analyzing health disparities in large spatiotemporal data sets: an example using breast cancer incidence data (Northern and Southern California, 1988--2002).

Efforts to monitor, investigate, and ultimately eliminate health disparities across racial/ethnic and socioeconomic groups can benefit greatly from spatiotemporal models that enable exploration of spatial and temporal variation in health. Hierarchical Bayes methods are well-established tools in the statistical literature for fitting such models, as they permit smoothing of unstable small-area rates. However, issues presented by 'real-life' surveillance data can be a barrier to routine use of these models by epidemiologists. These include (1) shifting of regional boundaries over time, (2) social inequalities in racial/ethnic residential segregation, which imply differential spatial structuring across different racial/ethnic groups, and (3) heavy computational burdens for large spatiotemporal data sets. Using data from a study of changing socioeconomic gradients in female breast cancer incidence in two population-based cancer registries covering the San Francisco Bay Area and Los Angeles County, CA (1988--2002), we illustrate a two-stage approach to modeling health disparities and census tract (CT) variation in incidence over time. In the first stage, we fit race- and year-specific spatial models using CT boundaries normalized to the U.S. Census 2000. In stage 2, temporal patterns in the race- and year-specific estimates of racial/ethnic and socioeconomic effects are explored using a variety of methods. Our approach provides a straightforward means of fitting spatiotemporal models in large data sets, while highlighting differences in spatial patterning across racial/ethnic population and across time.