Inference on dynamic models for non-Gaussian random fields using INLA

Abstract

Robust time series analysis is an important subject in statistical modeling. Models based on Gaussian distribution are sensitive to outliers, which may imply in a significant degradation in estimation performance as well as in prediction accuracy. State-space models, also referred as Dynamic Models, is a very useful way to describe the evolution of a time series variable through a structured latent evolution system. Integrated Nested Laplace Approximation (INLA) is a recent approach proposed to perform fast approximate Bayesian inference in Latent Gaussian Models which naturally comprises Dynamic Models. We present how to perform fast and accurate non-Gaussian dynamic modeling with INLA and show how these models can provide a more robust time series analysis when compared with standard dynamic models based on Gaussian distributions. We formalize the framework used to fit complex non-Gaussian space-state models using the R package INLA and illustrate our approach with a simulation study and a Brazilian homicide rate dataset.