B1098
Title: Similarity- and neighborhood-based dynamic models
Authors: Helena Baptista - Universidade Nova de Lisboa (Portugal) [presenting]
Abstract: Conditionally specified Gaussian Markov random field models with adjacency-based neighborhood weight matrix, have been the mainstream approach to spatial smoothing in Bayesian disease mapping. A conditionally specified Gaussian random field (GRF) model is proposed with a similarity-based non-spatial weight matrix to facilitate non-spatial smoothing. The model, named similarity-based GRF (with respect to the disease determinant factor), was motivated to model disease data in situations where the underlying small area relative risks do not vary systematically in space. More recently, the model proposed has proven to identify with greater accuracy high-risk areas in cases when the appropriate mix between local and global smoothing is not constant across the region. COVID-19 was the opportunity to explore the adequacy of the model to data from a contagious disease, very likely to be spatially and temporally positively correlated. On top of modelling, the model is used for predictions. The similarity-based GRF model shows a higher prediction power than the comparative models, proving to be an important tool for disease management and resource allocation. One important conclusion, there is no model that will always adequately explain and predict any phenomenon. Epidemics have waves and flexibility in the modelling process is key.
