Sampling Bias Worsen the Predictive Ability of Niche Models

Abstract

Purpose: Ecological Niche Modeling (ENM) and Species Distribution Modeling (SDM) have become powerful tools in biology, biogeography, paleoecology and biodiversity conservation. ENM and SDM are evaluated using metrics that take into account the errors and successes of the models in predicting the presence or absence of species in certain locations. Here, we evaluate the effects of sampling bias on the relationship between evaluative metrics and the predictive capacity of models.   Theoretical framework: ENM and SDM are powerful tools with extensive potential for use, but in order for them to produce useful results, they need to be constructed and validated appropriately. The occurrence data used in both processes may not have been collected randomly, which can lead to issues. An investigation into potential problems arising from the use of non-randomly collected occurrence data is necessary, as new issues may arise from simply filtering the data and reducing the number of occurrence records.   Material and Methods: We use Virtual Species (VS) to evaluate the effect of sampling bias. Using VS is the most robust approach for this type of testing, as we know the entire VS distribution.   Results and conclusion: Our results showed that sampling bias reduces the predictive capacity of the ENM and SDM models. We did not find a consistent pattern of the effect of sampling bias on the relationship between evaluation metrics and the predictive capacity of models. The effect size varied between different bias intensities. We emphasize that reducing the strength of the bias is one of the most efficient ways to minimize this problem.