A retrospective approach for evaluating ecological niche modeling transferability over time: the case of Mexican endemic rodents

Abstract

Ecological niche modeling (ENM) is a valuable tool for inferring suitable environmental conditions and estimating species’ geographic distributions. ENM is widely used to assess the potential effects of climate change on species distributions; however, the choice of modeling algorithm introduces substantial uncertainty, especially since future projections cannot be properly validated. In this study, we evaluated the performance of seven popular modeling algorithms—Bioclim, generalized additive models (GAM), generalized linear models (GLM), boosted regression trees (BRT), Maxent, random forest (RF), and support vector machine (SVM)—in transferring ENM across time, using Mexican endemic rodents as a model system. We used a retrospective approach, transferring models from the near past (1950–1979) to more recent conditions (1980–2009) and vice versa, to evaluate their performance in both forecasting and hindcasting. Consistent with previous studies, our results highlight that input data quality and algorithm choice significantly impact model accuracy, but most importantly, we found that algorithm performance varied between forecasting and hindcasting. While no single algorithm outperformed the others in both temporal directions, RF generally showed better performance for forecasting, while Maxent performed better in hindcasting, though it was more sensitive to small sample sizes. Bioclim consistently showed the lowest performance. These findings underscore that not all species or algorithms are suited for temporal projections. Therefore, we strongly recommend conducting a thorough evaluation of the data quality—in terms of quantity and potential biases—of the species of interest. Based on this assessment, appropriate algorithm(s) should be carefully selected and rigorously tested before proceeding with temporal transfers.