A comparison of multiple neighborhood matrix specifications for spatio-temporal model fitting: a case study on COVID-19 data

Establishing proper neighbor relations between a set of spatial units under analysis is essential when carrying out a spatial or spatio-temporal analysis. However, it is usual that researchers choose some of the most typical (and simple) neighborhood structures, such as the first-order contiguity matrix, without exploring other options. In this paper, we compare the performance of different neighborhood matrices in the context of modeling the weekly relative risk of COVID-19 over small areas located in or near Valencia, Spain. Specifically, we construct contiguity-based, distance-based, covariate-based (considering mobility flows and sociodemographic characteristics), and hybrid neighborhood matrices. We evaluate the goodness of fit, the overall predictive quality, the ability to detect high-risk spatio-temporal units, the capability to capture the spatio-temporal autocorrelation in the data, and the goodness of smoothing for a set of spatio-temporal models based on each of the neighborhood matrices. The results show that contiguity-based matrices, some of the distance-based matrices, and those based on sociodemographic characteristics perform better than the matrices based on k-nearest neighbors and those involving mobility flows. In addition, we test the linear combination of some of the constructed neighborhood matrices and the reweighting of these matrices after eliminating weak neighbor relations, without any model improvement.