(f) Governance and management covariates

We examined four covariates describing marine resource management and governance that might explain variation in social-ecological vulnerability to coral bleaching. MPAs and fisheries regulations have been promoted as tools to increase ecological recovery potential and reduce exposure to local stressors [44,65], although the empirical evidence is not conclusive [66]. General governance characteristics of an island, such as political stability, may influence socio-economic adaptive capacity and socio-economic sensitivity [67]. We assessed MPA coverage (as a percentage of each island's coral reef area; electronic supplementary material, table S17); fisheries regulations that control fishing effort or reduce fishing pressure on key species and/or life-history stages (electronic supplementary material, table S18); the World Bank's Worldwide Governance Indicator (WGI) score that reflects six dimensions of governance: voice and accountability, political stability and the absence of violence, government effectiveness, regulatory quality, rule of law, and control of corruption [68] (electronic supplementary material, table S19); and the sovereign status of each island (independent nation versus overseas territory, electronic supplementary material, table S19). Some of these covariates (e.g. WGI scores) have been used as indicators of adaptive capacity [24], but we chose to use more specific indicators with clear hypothesized or empirical links to adaptive capacity (see electronic supplementary material for explanations of each indicator chosen); our covariates represent variables that have been suggested by previously published research to be potentially important factors in explaining vulnerability more generally.

We developed generalized linear models (GLMs) for all components of vulnerability, modelling each component as a function of the four covariates. Model selection was performed in R [69] in stepwise fashion. We used the Akaike information criterion, squared Pearson's correlation coefficient (R²), and unadjusted squared deviance (D²) to select the most parsimonious and best-fit model. We also used Kruskal–Wallis tests to assess differences in the components of vulnerability, their constituent variables, and the covariates based on sovereign status while accounting for non-normal data distributions.