CFE 2019: Start Registration
View Submission - CMStatistics
Title: Statistical scalability of approximate inference for spatial models Authors:  Helen Ogden - University of Southampton (United Kingdom) [presenting]
Abstract: Even for simple normal models for spatial data, calculating the likelihood function can be infeasible for large datasets, as the cost of calculating the likelihood grows cubically with the number of data points. Because of this, many approximations to the likelihood have been proposed, all designed to be computationally scalable, so that the cost of computing the likelihood approximation does not grow too quickly with the size of the data. We study the statistical properties of inference with several families of approximate likelihoods, each of which involves tuning parameters which control a trade-off between computational cost and accuracy of the approximation. We discuss how the tuning parameters should be chosen to maintain good statistical properties as the amount of data grows, and discuss the implications of this for the scalability of the resulting inference.