Predictive performance of substrate, biota and lobster models was evaluated using the area-under-the-curve of receiver-operating characteristics (ROC) to determine the discriminatory ability of classification tree models.

Correct classification, sensitivity and specificity results were used to evaluate prediction accuracy, using Pfair as the threshold. Validation data consisted of 25% of dataset withheld from model development. The effect of spatial dependency on model accuracy was investigated using validation data that was spatially independent.