Bootstrap investigation of the stability of disease mapping of Bayesian cancer relative risk estimations.

BACKGROUND: Bayesian approaches to disease mapping of relative risks are useful for rare disease when geographical units have very different population sizes. As Bayesian approaches may induce very different estimations, it is useful to consider the stability of the estimations as a criterion for evaluating the quality of the results. MATERIAL: Cancer incidence data, from the Isere cancer registry (France) over the 1985-1994 period, have been used to check the proposed method: the study is based on 22 cancer sites among males and 24 among females.
METHOD: A bootstrap approach has been retained to evaluate the stability of the estimations. The coefficient of variation was chosen as an indicator of stability. Three Bayesian models corresponding to global, local and combined smoothing techniques, have been considered. The stability analysis has taken account of the results of spatial autocorrelation and heterogeneity tests.
RESULTS: Bayesian approaches do not necessarily lead to stable estimations. The local smoothing approach induces estimations that are often unstable. The global smoothing approach is the most stable, but is conservative. Combined smoothing appears to be a good compromise if significant spatial variations and heterogeneity of relative risks exist.
CONCLUSION: Bayesian estimations of relative risks may be very unstable. However, when results of spatial autocorrelation and heterogeneity tests are taken into account to choose between the different Bayesian approaches, instability becomes negligible.