Open-source biodiversity databases contain a large amount of species occurrence records, but these are often spatially biased, which affects the reliability of species distribution models based on these records. Sample bias correction techniques include data filtering at the cost of record numbers or require considerable additional sampling effort. However, independent data are rarely available and assessment of the correction technique must rely on performance metrics computed with subsets of the only available (biased) data, which may be misleading. Here we assess the extent to which an acknowledged sample bias correction technique is likely to improve models' ability to predict species distributions in the absence of independent data. We assessed the variation in model predictions induced by the correction and model stochasticity. We present an index of the effect of correction relative to model stochasticity, the Relative Overlap Index (ROI). We tested whether the ROI better represented the effect of correction than classic performance metrics and absolute overlap metrics using 64 vertebrate species and 21 virtual species with a generated sample bias. When based on absolute overlaps and cross-validation performance metrics, we found no effect of correction, except for cAUC. When considering its effect relative to model stochasticity, the effect of correction depended on the site and the species. Virtual species enabled us to verify that the correction actually improved distribution predictions and the biological relevance of the selected variables at the sites with a clear gradient of sample bias, and when species distribution predictors are not correlated with sample bias patterns.
Comment: 16 pages, 6 figures + Appendix