For species distribution models, species frequency is termedprevalence and prevalence in samples should be similar to natural speciesprevalence, for unbiased samples. However, modelers commonly adjust samplingprevalence, producing a modeling prevalence that has a different frequency ofoccurrences than sampling prevalence. The separate effects of (1) use ofsampling prevalence compared to adjusted modeling prevalence and(2) modifications necessary in thresholds, which convert continuousprobabilities to discrete presence or absence predictions, to account forprevalence, are unresolved issues. We examined effects of prevalence andthresholds and two types of pseudoabsences on model accuracy. Use of samplingprevalence produced similar models compared to use of adjusted modelingprevalences. Mean correlation between predicted probabilities of the least(0.33) and greatest modeling prevalence (0.83) was 0.86. Mean predictedprobability values increased with increasing prevalence; therefore, unlikeconstant thresholds, varying threshold to match prevalence values waseffective in holding true positive rate, true negative rate, and speciesprediction areas relatively constant for every modeling prevalence. The areaunder the curve (AUC) values appeared to be as informative as sensitivity andspecificity, when using surveyed pseudoabsences as absent cases, but when theentire study area was coded, AUC values reflected the area of predictedpresence as absent. Less frequent species had greater AUC values whenpseudoabsences represented the study background. Modeling prevalence had amild impact on species distribution models and accuracy assessment metricswhen threshold varied with prevalence. Misinterpretation of AUC values ispossible when AUC values are based on background absences, which correlatewith frequency of species.
展开▼
