International Day Of Women in Statistics and Data Science

In the 1970s, the conditionally formulated Gaussian Markov random field (GMRF), known as the conditional autoregressive (CAR) model, was introduced in line with Tobler’s first law of geography: "everything is related to everything else, but near things are more related than distant things." The CAR model uses W, the well-known adjacency matrix that encodes the neighbourhood structure of a spatial lattice (e.g., two areas are neighbours if they share a common border).
Since its introduction, the CAR model has undergone many adaptations, with numerous adaptive models proposed in the literature. However, almost all of these models (to the best of our knowledge, all except ours) still adhere to Tobler’s law. Yet, much of the data collected today – often aggregated at the areal level for confidentiality or other reasons – does not necessarily follow this law.
In this talk, we will show that any extra information representing causes of or correlated with the phenomenon of interest can be used to define a similarity structure, rather than relying solely on geographical neighbourhoods. Using simulated data, we illustrate that similarity-based structures can be more effective than traditional neighbourhood-based structures for smoothing both local and global risks. We will show that the correct identification of high- and low-risk areas, crucial for public health planning and resource allocation, is better achieved when the similarity-based structure is used.