Efficient Bayesian Filtering and Inference for Integro-Difference Equation Models

Abstract

Integro-Difference Equation (IDE) models are a class of dynamic spatio-temporal models which describe the temporal evolution of a spatial process, usually to model advection and diffusion of some random field over time. While several papers have focused on inference and performance of these models, little attention has been given to computational comparisons of various methods of computing posterior densities and performing inference. We address this problem in two parts. Firstly, in inference for these models, Kalman filtering (and smoothing) is often used; either to compute marginal likelihoods, or to sample from the hidden process conditional on the rest of the parameters of the model. There are many variations of the Kalman filter and smoother with substantially varying performance profiles and stability: the standard Kalman filter, its square-root form, its information form, to name a few. The information filters, when set up correctly, can compute the likelihood in $\mathcal O(nr^2 + r^3)$ time per time point ($n$ being the number of observations per time point), where the computation in $n$ can be done in parallel, on models with independent observation error. In this paper, we analyse these forms in the context of IDEs, considering computational speed and stability. Secondly, there are two main methods of inference used for the kind of dynamical system that IDEs can be reduced to; directly using the marginal likelihood without the hidden state, or constructing a block Gibbs sampler with the hidden state. We compare these two approaches in terms of computational efficiency and practical performance in the IDE setting. Using a block Gibbs sampler for inference on the system parameters proves to be more efficient in most cases, with a specific real-data example providing an exception.