Similarity- and neighborhood-based dynamic models for infection data: Uncovering the complexities of the COVID-19 infection risks

According to the WHO, COVID-19, like any other contagious respiratory disease, spreads across populations between people who are in close contact with each other. By analysing weekly COVID-19 infection case data collected from the disease’s first reported case on 3 March 2020 to 22 April 2021 for two hundred seventy-eight local municipalities of Portugal, we show that the complexities of that spread vary, depend- ing on the level of cases. Accounting for spatial risk dependence and correlation is especially important when studying contagious diseases. This research uses a condi- tionally specified Gaussian random field model with a novel approach to characterize COVID-19 infection risk dependencies via similarities of areal-level covariate(s), defined within a Bayesian hierarchical model framework. Three main periods of the pandemic are studied, with different levels of the daily number of cases, a low period, a rising level period and the period with the highest level of cases. We show that the traditional conditional autoregressive model, defined by an adjacency-based neigh- bourhood matrix, may not adequately characterize the complex nature of infection risk dependence over space. We also illustrate that infection risk prediction and infec- tion case forecasting may help informing decisions that have difficulty-to-measure impacts and potentially save lives.