Reproducibility in species distribution modelling: only 19% of studies reproduced

Ferreira and Mensah attempted to reproduce the headline result of 218 species distribution modelling studies published between 2015 and 2024, using only what the authors had made available. Complete reproduction succeeded for 41 studies. A further 66 reproduced partially: the direction of the primary effect, but not its magnitude.

The failure modes are mundane and therefore fixable: unavailable occurrence data (47%), undocumented preprocessing (31%), and unspecified software versions (22%).

The finding that should worry editors most concerns data-availability statements. Studies in journals that mandate them were 2.8 times more likely to reproduce, but of the articles stating that data were available on request, 39% produced nothing when asked.

The authors propose a minimal reporting checklist, and demonstrate that it recovers reproducibility in 84% of a held-out validation set. It is published open access.