Evaluating the performance of spatio-temporal Bayesian models in disease mapping

摘要

In the last few decades there has been an improvement in the statistical methods used to display the geographical patterns of mortality and disease incidence. These methods consider spatial models to smooth the classical standardized mortality ratio (SMR). Nowadays, interest relies on extending these spatial models to incorporate time trends and spatio-temporal interactions due to the availability of historical high quality mortality registers recorded during the last 20 years. In this work, alternative Bayesian spatio-temporal models are fitted using MCMC techniques. The performance of these models is analyzed through a simulation study with two objectives in mind: the first one is to evaluate the relative bias and relative standard error of the posterior mean relative risks, and the second one is to investigate recent Bayesian decision rules to detect raised-risk areas in a spatio-temporal context. The simulation study is based on mortality data due to colbrectal cancer in males from Navarra, Spain, corresponding to four 5-year time windows. When there are a number of high-risk areas in some of the time periods we conclude that the bias of the posterior mean relative risks could be substantial. The decision rules to detect these high-risk areas should be determined with caution. A final rule combining alternative threshold and cutoff values for the different time periods seems to be needed.