Bovo E., Girardi P., Guzzinati S., Baracco M., Dal Cin A., Carpin E., Memo L., Martin G., Fiore A., Rugge M.

GRELL Ascension Meeting

Trento, 16 -18 may 2018



Laplace approximation methods as INLA (Integrated Nested Laplace Approximation) are being widely used in Bayesian inference, especially in the assessment of the spatial risk distribution as the Besag-York-Mollie (BYM) model. Compared to Monte Carlo simulations based on Markov Chains (MCMC), INLA is time saving; however, it can produce different estimates(Carroll 2015; De Smedt 2015). The aim of this study is to compare estimates produced by INLA with those computed with MCMC simulations.

We considered all cases of malignant cancers occurring in the year 2013 in Veneto Region (about five million of inhabitants). The area covered by the Veneto Cancer Registry includes the 96% of the population (n=556 municipalities). In our comparison, we considered 2 different primitive cancer sites: lung (males) and cervix cancer (females). The Standardized Incidence Ratios (SIRs) have been estimated by means of a BYM model, using the registry pool as a reference. SIR has been estimated by INLA and MCMC using R-INLA and R2WinBUGS packages on R environment. The differences were analyzed by varying the distribution of the spatial components of the model.

Estimates produced by INLA are comparable to MCMC, however they are affected by the distribution of the a-priori variable and by the number of observations. In the presence of non-informative a-priori distributions, the accuracy of the spatial parameter estimates parameter in INLA is reduced, while WinBUGS seems to be more robust. The performances of the two methods differ according to the number of observations: in fact, we observed wider differences in the SIR for cervix cancer, which is less frequent than lung cancer.

INLA is a user-friendly, fast and efficient method for spatial estimates. However, to avoid an over-smoothing of risks and/or an excessive imprecision of estimates, particular care is needed in the choice of the a-priori distribution parameter.