2.7. Classification and regression tree
A classification and regression tree (CART) approach was used to determine which traits were most likely to be lost from the matrix category of the forest–agriculture gradient as well as fragmented forests. For the forest–agriculture gradient, we identified species present in interior forests and absent from all other forest classifications. For fragmentation, we identified all species present in the continuous forests but absent from the forest fragments. The CART analysis recursively partitioned the predictor variables (traits) to explain the variability in the response variable, which was whether each species was present in at least one of the unmodified habitats, but none of the modified habitats. This applied to both types of disturbance regimes. We used the sum of squares around group means to establish splitting criteria, selected optimal tree size by K-fold cross-validation, and evaluated model fit with Pearson correlation of predicted to observed species loss (De’ath and Fabricius, 2000).