sabinaMBM: An R package for Multiscale Bayesian species distribution Modelling using INLA

Abstract

1. Regional species distribution models (SDM) calibrated over spatially restricted extents tend to truncate species' ecological niches. Existing nested SDM workflows integrate multi-scale information through sequential combination, which limits formal uncertainty propagation across scales and prevents regional predictions from being explicitly constrained within globally-informed niche boundaries. 2. We introduce sabinaMBM, an R package implementing joint multiscale Bayesian SDM within a unified probabilistic framework built on inlabru/R-INLA. It propagates uncertainty across scales without the computational bottlenecks of MCMC-based approaches. The framework offers multiple coupling architectures ranging from complete independence to hierarchical constraint that can be configured independently for intercepts and covariates. 3. In a range-margin population, hierarchical constraint most improves out-of-sample discrimination where regional data were scarcest, while leaving predictions unchanged where they already suffice, delivering gains precisely where sequential approaches are expected to struggle most. Applied to Quercus petraea across its Iberian trailing-edge, including a spatial field produced the largest single performance gain, consistent across every coupling configuration, and covariate responses diverged by scale for at least one climatic predictor. Under future climate, the constrained model yielded lower suitable habitat estimates and redistributed uncertainty in proportion to cross-scale agreement rather than uniformly. 4. sabinaMBM makes multiscale Bayesian inference accessible without specialist programming. This framework provides robust value for trailing-edge populations and spatial (invasive species) or temporal (climate change) projections where niche truncation risks ecologically implausible outcomes, while simultaneously delivering fine-resolution predictions with properly propagated uncertainty whenever global and regional covariates offer complementary information.