Global Sensitivity Analysis of Societal Resilience Using Shapley Values and Polynomial Chaos Expansion

ABSTRACT

Societies worldwide face increasingly complex and interconnected crises that challenge their capacity for resilience. Assessing which structural indicators are most strongly associated with resilience scores requires quantitative methods capable of handling interdependencies, nonlinearities, and limited sample sizes. This study applies established global sensitivity analysis tools within an empirical resilience setting characterized by correlated structural indicators and small cross‐national datasets. Using 124 indicators from the Joint Research Centre Resilience Dashboards, we analyze model‐based structural correlates of societal resilience across EU Member States using the Lloyd's Register Foundation World Risk Poll Resilience Index for 2021 and 2023. Following univariate Pearson correlation screening and Random Forest stability selection, Shapley‐based variance decomposition is computed on a Polynomial Chaos Expansion (PCE) surrogate, enabling equitable attribution under correlated inputs. For 2023, the results indicate a concentrated attribution structure, with a small core of indicators accounting for most explained variance ( R ² = 0.85). Active citizenship has the largest model‐based contribution, followed by antimicrobial resistance and years of life lost due to PM ₂ . ₅ exposure, suggesting strong associations with social capital and environmental health under sustained polycrisis conditions. Environmental innovation, ecological conditions, and macroeconomic buffering form a secondary tier of contributors. By contrast, the 2021 model exhibits a more diffuse attribution profile ( R ² = 0.82), with resilience scores associated more evenly with active citizenship, adult competences, digital readiness, and labor‐market adjustment. Leave‐one‐out validation suggests that these attribution patterns are not dominated by single‐country exclusions, although uncertainty remains substantial in a small cross‐sectional sample. Overall, the comparison indicates a shift in the fitted attribution structure from broader recovery‐oriented human‐capital and digital indicators in 2021 toward more concentrated structural contributions linked to social and environmental conditions in 2023, offering a transparent methodological framework for comparative resilience analysis in small‐sample policy settings.