Calculations of model prediction error on the training data indicated little or no improvement in model performance when more predictor variables are included, i.e. models with all predictor variables (full), or built stepwise (step).
In Table 3, row "p" represented the model performance with complete predictor variables, while row "pi = 0" represented that with variable "pi" removed.
Second, when the GLM analyses were repeated but with tree mortality included as a potential predictor, model performance improved (explaining 42 91% of the variation in edge parameters) and tree mortality was a significant predictor for all edge variables except liana abundance (Table S2).
The second choice, the partially customized (PC) score model, was built to evaluate differences in model performance when the same predictor variables were used.
The diminishing return on model performance when adding statistically significant predictors to the model was also noted in the model building process of DPoRT and was one of the reasons that DPoRT maintained good discrimination, even with considerable constraints on variable selection (Rosella et al. 2010).
For catchments with low moisture homogeneity (IM < 80%), IM is a better predictor of model performance improvement than ITV; whereas for catchments with high moisture homogeneity (IM > 80%), ITV is a better predictor of performance improvement than IM.
Overall, the contribution of soil map derived predictors to the model performance was rather weak.
Migration beyond the western park boundary was less influenced by these predictors and process model performance suggested an important control on recent migrations was excluded.
For measures of pragmatic model performance, a strategy for handling missing predictors in the future clinical setting must be chosen and is evaluated as part of the risk prediction algorithm.
Predictions P7 P9 (which reduce to P1 P3 for individuals with fully observed covariates) may be used to estimate measures of pragmatic model performance, under the assumption that missing predictors in the future clinical setting will be handled using multiply imputation in the same way.
