Multi-Level Spatial Comparative Judgement Models To Map Deprivation

While current comparative judgement models provide strong algorithmic efficiency, they remain data inefficient, often requiring days or weeks of extensive data collection to provide sufficient pair- wise comparisons for stable and accurate parameter estimation. This disparity between data and algorithm efficiency is preventing widespread adoption, especially so in challenging data-collection environments such as mapping human rights abuses. We address the data inefficiency challenge by introducing the finite element Gaussian process Bradley–Terry mixture model, an approach that significantly reduces the number of pairwise comparisons required by comparative judgement mod- els. This is achieved via integration of prior spatial assumptions, encoded as a mixture of functions, each function introducing a spatial smoothness constraint at a specific resolution. These functions are modelled nonparametrically, through Gaussian process prior distributions. We use our method to map deprivation in the city of Dar es Salaam, Tanzania and locate slums in the city where poverty reduction measures can be carried out.