Global data gaps contribute to low predictive accuracy in human pressure-based biodiversity models

Amidst the ongoing biodiversity crisis, there is high demand for spatially explicit biodiversity indicators that can support conservation planning and national reporting. Global models that estimate the impacts of human pressures on biodiversity provide crucial insights, but their use in spatial projections calls for more systematic evaluation of how accurately they can predict biodiversity patterns at fine spatial scales. This is especially important because spatial projections require models to make predictions under a wide range of environmental and geographic contexts. Here we evaluate the generality of two different pressure-response models for estimating alpha and beta diversity, relative to ecologically intact reference sites, using a global dataset of 25,987 sites from 681 biodiversity studies. Generality is operationalized as the model accuracy when making out-of-sample predictions in sampled populations (generalizability) as well as in other contexts (transferability).

We find that mixed models with study-level random effects – commonly used in meta-analyses and forming the basis of several biodiversity indicators – exhibit generally low site-level accuracies. This reflects dependence on a limited set of averaged fixed effects and strong attribution of variation to the random effects, which cannot be used out-of-sample. In comparison, a model structure that incorporates biogeographic–taxonomic attributes together with environmental covariates achieves higher accuracy within contexts represented in training data. However, accuracy is low when predicting into new contexts, due to distribution shifts between training and test data. These patterns hold for both site-level diversity measures and for differences between paired sites.

Although both models estimate consistent and reasonable responses to land use, the results illustrate a large gap between effect-size inference and spatially explicit prediction. Models are essential for informed conservation efforts, but their applicability is fundamentally constrained by the availability and distribution of underlying data. Whereas countries with extensive data can build high-fidelity national indicators, accelerated data collection and macroecological model development are needed to better support data-poor regions with actionable biodiversity insights.