Cluster-specific ranking and variable importance for Scottish regional deprivation via vine mixtures

TL;DR

This paper introduces a vine mixture model approach to cluster Scottish regions based on deprivation indicators, enabling a nuanced ranking and variable importance assessment in an unsupervised context.

Contribution

It develops a novel vine mixture framework for clustering and ranking regions using socioeconomic data, with a new method for variable importance in unsupervised learning.

Findings

Income and employment are key deprivation drivers.

Health and crime indicators are less influential.

Clusters reveal socioeconomic disparities across Scottish zones.

Abstract

Socioeconomic deprivation is a key determinant of public health, as highlighted by the Scottish Government's Scottish Index of Multiple Deprivation (SIMD). We propose an approach for clustering Scottish zones based on multiple deprivation indicators using vine mixture models. This framework uses the flexibility of vine copulas to capture tail dependent and asymmetric relationships among the indicators. From the fitted vine mixture model, we obtain posterior probabilities for each zone's membership in clusters. This allows the construction of a cluster-driven deprivation ranking by sorting zones according to their probability of belonging to the most deprived cluster. To assess variable importance in this unsupervised learning setting, we adopt a leave-one-variable-out procedure by refitting the model without each variable and calculating the resulting change in the Bayesian information criterion. Our analysis of 21 continuous indicators across 1964 zones in Glasgow and the surrounding areas in Scotland shows that socioeconomic measures, particularly income and employment rates, are major drivers of deprivation, while certain health- and crime-related indicators appear less influential. These findings are consistent across the approach of variable importance and the analysis of the fitted vine structures of the identified clusters.