Deciphering ecology from statistical artefacts: Competing influence of sample size, prevalence and habitat specialization on species distribution models and how small evaluation datasets can inflate metrics of performance

摘要: AIM: Sample size and species characteristics, including prevalence habitat specialization, can influence the predictive performance of distribution models (SDMs). There is little agreement, however, on which metric model to use. Here, we directly compare AUC partial ROC as metrics SDM through analyses effects traits sample performance. LOCATION: Three counties dominated by agricultural lands coniferous forest in Oregon's Willamette Valley Coast Range ecoregions. METHODS: We systematically reduced a large avian point count dataset alter sizes 22 songbird. used boosted regression trees run SDMs for each species, quantified mixed size, prevalence, specialization performance, calculated ROC, across species. with subset independent evaluation data separately more comprehensively investigate differences metrics. RESULTS: found positive quadratic effect strongly both weak no AUC. Contrary expectations, when evaluated data, was consistently highest smallest sizes. These small had correspondingly datasets. Partial or showed expected correlation between MAIN CONCLUSIONS: that datasets artificially inflate ROC. With literature recommended minimum low three, attention must be given