Two-stage MCMC for Fast Bayesian Inference of Large Spatio-temporal Ordinal Data, with Application to US Drought

TL;DR

This paper introduces a two-stage MCMC algorithm for efficient Bayesian inference on large-scale spatio-temporal ordinal data, effectively capturing dependencies without restrictive simplifications, demonstrated on US drought data.

Contribution

The paper presents a novel two-stage MCMC method that enables fast, accurate Bayesian inference for large spatio-temporal ordinal datasets, overcoming computational challenges of traditional models.

Findings

Significant computational speed-up over single-stage models

Posterior distributions closely match more costly methods

Successful application to large US drought dataset

Abstract

High dimensional space-time data pose known computational challenges when fitting spatio-temporal models. Such data show dependence across several dimensions of space as well as in time, and can easily involve hundreds of thousands of observations. Many spatio-temporal models result in a dependence structure across all observations and can be fit only at a substantial computational cost, arising from dense matrix inversion, high dimensional parameter spaces, poor mixing in Markov Chain Monte Carlo, or the impossibility of utilizing parallel computing due to a lack of independence anywhere in the model fitting process. These computational challenges are exacerbated when the response variable is ordinal, and especially as the number of ordered categories grows. Some spatio-temporal models achieve computational feasibility for large datasets but only through overly restrictive model simplifications, which we seek to avoid here. In this paper we demonstrate a two-stage algorithm to fit a Bayesian spatio-temporal model to large datasets when the response variable is ordinal. The first stage models locations independently in space, capturing temporal dependence, and can be run in parallel. The second stage resamples from the first stage posterior distributions with an acceptance probability computed to impose spatial dependence from the full spatio-temporal model. The result is fast Bayesian inference which samples from the full spatio-temporal posterior and is computationally feasible even for large datasets. We quantify the substantial computational gains our approach achieves, and demonstrate the preservation of the posterior distribution as compared to the more costly single-stage model fit. We apply our approach to a large spatio-temporal drought dataset in the United States, a dataset too large for many existing spatio-temporal methods.