Bayesian Gaussian models for interpolating large-dimensional data at misaligned areal units

TL;DR

This paper introduces a Bayesian Gaussian model with novel basis functions for interpolating large, sparse, and misaligned areal spatial data, demonstrated through real and simulated examples, improving spatial prediction accuracy.

Contribution

It develops a new basis function approach within a Bayesian Gaussian framework to handle large, irregular, and misaligned areal data for spatial prediction.

Findings

Effective in handling large-dimensional spatial data

Improves prediction at misaligned geographical units

Validated through real and simulated data

Abstract

Areal level spatial data are often large, sparse and may appear with geographical shapes that are regular or irregular (e.g., postcode). Moreover, sometimes it is important to obtain predictive inference in regular or irregular areal shapes that is misaligned with the observed spatial areal geographical boundary. For example, in a survey the respondents were asked about their postcode, however for policy making purposes, researchers are often interested to obtain information at the SA2. The statistical challenge is to obtain spatial prediction at the SA2s, where the SA2s may have overlapped geographical boundaries with postcodes. The study is motivated by a practical survey data obtained from the Australian National University (ANU) Poll. Here the main research question is to understand respondents' satisfaction level with the way Australia is heading. The data are observed at 1,944 postcodes among the 2,516 available postcodes across Australia, and prediction is obtained at the 2,196 SA2s. The proposed method also explored through a grid-based simulation study, where data have been observed in a regular grid and spatial prediction has been done in a regular grid that has a misaligned geographical boundary with the first regular grid-set. The real-life example with ANU Poll data addresses the situation of irregular geographical boundaries that are misaligned, i.e., model fitted with postcode data and hence obtained prediction at the SA2. A comparison study is also performed to validate the proposed method. In this paper, a Gaussian model is constructed under Bayesian hierarchy. The novelty lies in the development of the basis function that can address spatial sparsity and localised spatial structure. It can also address the large-dimensional spatial data modelling problem by constructing knot based reduced-dimensional basis functions.