Model averaging fails to improve the extrapolation capability of the island species-area relationship

Aim One of the main applications of the island species-area relationship (SAR) is to predict species richness in areas of habitat too large to be sampled, but there are few clear guidelines for choosing an appropriate model for this purpose. We therefore aimed to test whether a multi-model averaging approach could improve the accuracy of predictions made by extrapolating the ISAR. Specifically, we compared the performance of multi-model averaging with that of the default ISAR model of choice, the power model, in predicting species richness in large habitat islands. Location Global Taxa Vertebrates, invertebrates and plants Methods We removed the largest islands from 120 habitat island datasets, and fitted both the power model and a multi-model average curve (averaging the predictions of up to 20 ISAR models) to this filtered dataset. We then assessed the accuracy of both approaches in predicting the species richness of the largest island in the original dataset using the log error of extrapolation (LEE) metric. A generalized additive regression modelling framework was used to determine whether any dataset characteristics could explain variation in the LEE values for the power model (LEE-POW). Results The power model gave the more accurate richness predictions for 58% of the analysed datasets and the multi-model averaged curve gave the more accurate predictions for the remaining 42%. Both the power models (61% of LEE-POW values were positive) and the multi-model averaged curve (60% were positive) had a slightly greater tendency to over predict the observed richness. The confidence intervals were also on average narrower for the power model predictions (median 95% confidence interval width = 18 species) than for the multi-model averaged curve predictions (median 95% confidence interval width = 78). The range in island areas and richness values explained a small amount of the variation in LEE-POW. Main conclusions Contrary to expectation, multi-model averaging was less accurate than the power model in the majority of cases, and thus does not appear to be a panacea for uncertainty in model choice when extrapolating the ISAR. However, further research is urgently needed to evaluate the performance of a multi-model averaging approach at larger spatial scales.