Visualizing and testing the impact of place on late-stage breast cancer incidence: A non-parametric geostatistical approach

Abstract

This paper describes the combination of three-way contingency tables and geostatistics to visualize the non-linear impact of two putative covariates on individual-level health outcomes and test the significance of this impact, accounting for the pattern of spatial correlation and correcting for multiple testing. The methodology is used to explore the influence of distance to mammography clinics and census-tract poverty level on the rate of late-stage breast cancer diagnosis in three Michigan counties. Incidence rates are significantly lower than the area-wide mean (18.04%) mainly in affluent neighbourhoods [0–5% poverty], while higher incidences are mainly controlled by distance to clinics. The new simulation-based multiple testing correction is very flexible and less conservative than the traditional false discovery rate approach that results in a majority of tests becoming non-significant. Classes with significantly higher frequency of late-stage diagnosis often translate into geographic clusters that are not detected by the spatial scan statistic.