Impact of socioeconomic inequalities on geographic disparities in cancer incidence: comparison of methods for spatial disease mapping. Issue 1 (December 2016)

• Record Type: Journal Article
• Authors: Goungounga, Juste; Gaudart, Jean; Colonna, Marc; Giorgi, Roch
• Is Part Of: BMC medical research methodology. Volume 16:Issue 1(2016)
• Page Start: 1
• Page End: 14
• Publication Date: 2016-12
• Subjects: Spatial analysis -- Cluster detection -- Cancer -- Oblique decision tree; Medicine -- Research -- Methodology -- Periodicals; 610.72
• DOI: 10.1186/s12874-016-0228-x
• Languages: English
• ISSNs: 1471-2288

Abstract

Abstract Background The reliability of spatial statistics is often put into question because real spatial variations may not be found, especially in heterogeneous areas. Our objective was to compare empirically different cluster detection methods. We assessed their ability to find spatial clusters of cancer cases and evaluated the impact of the socioeconomic status (e.g., the Townsend index) on cancer incidence. Methods Moran's I, the empirical Bayes index (EBI), and Potthoff-Whittinghill test were used to investigate the general clustering. The local cluster detection methods were: i) the spatial oblique decision tree (SpODT); ii) the spatial scan statistic of Kulldorff (SaTScan); and, iii) the hierarchical Bayesian spatial modeling (HBSM) in a univariate and multivariate setting. These methods were used with and without introducing the Townsend index of socioeconomic deprivation known to be related to the distribution of cancer incidence. Incidence data stemmed from the Cancer Registry of Isère and were limited to prostate, lung, colon-rectum, and bladder cancers diagnosed between 1999 and 2007 in men only. Results The study found a spatial heterogeneity (p < 0.01) and an autocorrelation for prostate (EBI = 0.02;p = 0.001), lung (EBI = 0.01;p = 0.019) and bladder (EBI = 0.007;p = 0.05) cancers. After introduction of the Townsend index, SaTScan failed in finding cancers clusters. This introduction changed the results obtained with the other methods. SpODT identified five spatial classes (p < 0.05): four in the Western and one in the Northern parts of the study area (standardized incidence ratios: 1.68, 1.39, 1.14, 1.12, and 1.16, respectively). In the univariate setting, the Bayesian smoothing method found the same clusters as the two other methods (RR >1.2). The multivariate HBSM found a spatial correlation between lung and bladder cancers (r = 0.6). Conclusions In spatial analysis of cancer incidence, SpODT and HBSM may be used not only for cluster detection but also for searching for confounding or etiological factors in small areas. Moreover, the multivariate HBSM offers a flexible and meaningful modeling of spatial variations; it shows plausible previously unknown associations between various cancers.