Actionable inference for biodiversity change hinges on representative data and model design

Jakob Nyström , Jeffrey R. Smith, Lisa Mandle, Andrew Gonzalez , Thomas B. Schön, Tobias Andermann 

Amidst the global biodiversity crisis, there is high demand for spatially explicit biodiversity indicators. Global models that quantify impacts of human pressures provide important insights for conservation, but their performance in spatial projections has not been systematically tested. W evaluate this using PREDICTS data, finding that, despite land-use impacts in line with previous research, there is a challenging gap between effect size inference and prediction. We find that mixed models with study attributes as random effects - common in biodiversity meta-analysis and indicators - exhibit low predictive accuracy, driven by reliance on highly averaged fixed effects. Ecologically structured models that replace study random effects with biome, realm, and taxonomic parameters, show improved but still modest results in sampled contexts. Yet, performance when extending predictions to other contexts remains low, due to distribution shifts in environmental factors and conditional responses. Our results highlight tensions between high-resolution predictor availability and the granularity at which responses can be reliably estimated. While models are essential for informed conservation efforts, their applicability is fundamentally constrained by data availability. Countries with extensive data can build higher-fidelity national indicators, but accelerated and systematically structured data collection is needed to support data-poor regions with localized, actionable biodiversity insights.