Incorporating sampling bias into ENM/SDM permutation tests: new methods and a case study on neotropical ants

Spatial sampling bias in occurrence data can generate spurious environmental associations in models of species distributions and ecological niches, and can also undermine inferences based on permutation tests made using these models. Geographic randomization tests are often used to generate distributions of expected behavior under the null hypothesis that species occurrences are unrelated to environmental predictors. Here we quantify how ignoring spatial sampling bias in these tests affects Type I error (i.e. false positives) for hypothesis tests and distorts estimates of methodological bias, and we evaluate a simple remedy that samples pseudoabsences in proportion to empirical bias patterns. We used a virtual species approach to simulate scenarios where no ecological processes were operating but where sampling was spatially biased to represent a worst‐case scenario. We evaluated the effects of ignoring this bias on Type I error rates for multiple tests of model significance, tests of niche similarity, and tests for measuring methodological bias in model transfer scenarios. When randomization tests were conducted without accounting for sampling bias, Type I error was substantially inflated for hypothesis tests. Incorporating the correct spatial sampling bias model into randomization analyses reduced error rates for all randomization tests, and in most cases brought them down to approximately nominal levels. Analyses of empirical data for a group of South American ant species produced qualitatively similar but generally attenuated effects, but in some cases test results became significant only after bias was accounted for. We conclude that incorporating empirical spatial sampling bias estimates into geographic randomization tests substantially mitigates inflated Type I error and can materially alter estimates of methodological bias. We recommend including spatial sampling bias estimates in randomization tests where possible, and estimates of spatial autocorrelation where reliable bias estimates are not practical.