Figure 3.

Random forest variable importance measures in (A) the full and (B) reduced model. A variable’s importance is measured by the difference of the model’s prediction accuracy (=relative frequency of correct classifications) before and after random permutation of a variable’s values. The permutation is used to nullify any relations of a variable to the outcome or other variables. The prediction accuracy stays unaffected by permutation of a variable if it is not of relevance for prediction. The variable importance takes a value of zero (or small negative values resulting from random variation) is such a case.