Reproducibility of published species distribution models: a systematic reanalysis of 218 studies

Abstract

Species distribution models inform conservation policy, yet the reproducibility of published models has not been assessed at scale. We attempted to reproduce the headline results of 218 species distribution modelling studies published between 2015 and 2024, using only the data and code the authors made available. Complete reproduction was possible for 41 studies (18.8%); partial reproduction, defined as recovering the direction but not the magnitude of the primary effect, was possible for a further 66 (30.3%). Failure was attributable to unavailable occurrence data in 47% of cases, undocumented preprocessing in 31%, and unspecified software versions in 22%. Studies published in journals with mandatory data-availability statements were 2.8 times more likely to be reproducible, but statements were frequently unfulfilled: 39% of articles declaring data availability on request produced no data when asked. We propose a minimal reporting checklist and demonstrate that it recovers reproducibility in 84% of a held-out validation set.