Assessment of the INLA approach on gerarchic bayesian models for the spatial disease distribution: a real data application

The use of approximate methods as the INLA (Integrated Nested Laplace Approximation) approach is being widely used in Bayesian inference, especially in spatial risk model estimation where the Besag-York-Mollie (BYM) model ` has found a proper use. INLA appears time saving compared to Monte Carlo simulations based on Markov Chains (MCMC), but it produces some differences in estimates [1, 2]. Data from the Veneto Cancer Registry has been considered with the scope to compare cancer incidence estimates with INLA method and with two other procedures based on MCMC simulation, WinBUGS and CARBayes, under R environment. It is noteworthy that INLA returns estimates comparable to both MCMC procedures, but it appears sensitive to the a-priori distribution. INLA is fast and efficient in particular with samples of moderate-high size. However, care must to be paid to the choice of the parameter relating to the a-priori distribution.