AN EMPIRICAL COMPARISON OF MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION METHODS FOR MULTIVARIATE DISEASE MAPPING

摘要

There are several publications comparing estimating methods for fitting a univariate spatial model. Such comparisons are rare for multivariate spatial models, which have been the focus of recent research in epidemiology. Using mortality data from the Greater Glasgow Health Board, we compare a full Bayesian approach with a frequentist approach for estimating relative risks of two diseases simultaneously. The regression parameters are estimated similarly under the two methods. Consistent with previous comparisons for univariate disease mapping, variance estimates are generally attenuated under a frequentist method. While a frequentist approach has computational limitations, a Bayesian approach, through a Markov Chain Monte Carlo (MCMC) algorithm, is less cumbersome to implement. In addition, a full Bayesian approach allows complex and flexible hierarchical modelling, and it provides more reliable estimates and predictions for many realistic epidemiological problems. The advent of the freely available WinBUGS package and the inclusion of MCMC methods in packages such as MLwiN should make Bayes methods easily accessible to most epidemiologists tornincorporate prior information about the unknown parameters into their models.