Figure 3. Summary of variable importance values from a random forest model fitted to the discovery dataset.

Included variables are listed on the vertical axis, with corresponding variable importance for each on the horizontal axis, Importance is determined using “conditional permutation accuracy”, calculated as a the average difference in model accuracy between the fitted model and alternative versions obtained via random permutations of the variable values. Variables are considered significant predictors in the random forest if their variable importance value is above the absolute value of the lowest negative-scoring variable [43].