Results
The number of diurnal landbird species that disappeared from each island ranged from zero (Malawali and Mantanani)to 15 (Bintan) (Table 1). The percentage of forest-dwelling birds that disappeared from the islands ranged from 0% (Balambangan, Malawali, and Mantanani)to 100% (Mengalum; Table 1). Sampling saturation was reached for all islands except Mantanani (Fig. 2), and observed species richness for all islands was very close to themean estimated richness (88.7–97.9%; Table 1). Some species vanished from all the islands where they previously occurred (Scarlet-rumped Trogon), whereas others disappeared from some but not all islands (Oriental Pied
Hornbill [Anthracoceros albirostris]) (Supporting Information).
The regression model of percentage of forest cover explained 34.8% of variation in forest-bird extinctions across seven islands (Fig. 3). Extinction rate was lower when forest cover was high (more habitat). Nevertheless, statistical support for the regression was weak, probably
because our small sample size was small (seven islands). The relative AICc weight (sums to 1.0) was strongest (0.825) for the null (no predictor) model. That is, the model in which extirpation rate and forest cover were not related was 4.7 times better supported by the data (on the basis of the evidence ratio of the AICc weights)even though the correlation with forest cover was reasonably strong. We considered this an issue of data paucity rather than structural inadequacy (i.e., data from additional islands would be needed to reinforce the strength of evidence for the predictive relationships, but relevant historical data are lacking). When only forest bird species were modelled, the variation explained dropped slightly to 32%.
For the classification-tree trait comparison between extinct and extant birds, all predictors except forest specialization and primary diet were selected. The full tree had 23 terminal nodes. The final and optimum (parsimonious)tree from our classification tree analysis, however,
had five terminal nodes (Fig. 4). Overall, the classification models were highly accurate (>84%) and statistically significant (Cohen’s kappa p < 0.001) (Table 2).
The traits retained in the optimal tree were Indomalayan realm endemism, minimum clutch size, habitat breadth, and mass (in decreasing order of importance) (Fig. 4). These same four traits were also the most important in predicting local extinction of Bintan birds on the basis
of random forests. In particular, predictors that resulted in the largest mean decrease in accuracy of the classification were the most important: Indomalayan realm endemism (2.99), minimum clutch size (2.56), habitat breadth (1.94), and mass (1.67). Overall, support for the
classification-tree model was corroborated by the random forest model (Table 2).The probability of extinction in Bintan birds varied from 0% to 100%, depending on the combination of these four traits (Fig. 4). For instance, this model estimated a 100% probability of extinction (three species) for the most vulnerable collection of traits. Bird species in Bintan that are endemic to the Indomalayan realm, have a minimum clutch size of at least 1.5, have a habitat breadth
of <1.5, and are >25.9 g (second terminal node from the right; Fig. 4). In addition, birds that are endemic to the Indomalayan realm and have smaller minimum clutch sizes had a 75% extinction probability (first terminal node from the right; Fig. 4).
การแปล กรุณารอสักครู่..