Caret varImp for randomForest model

Question

I'm having trouble understanding how the varImp function works for a randomForest model with the caret package. In the example below feature var3 gets zero importance using caret's varImp function, but the underlying randomForest final model has non-zero importance for feature var3. Why is this the case?

require(randomForest)
require(caret)


rf <- train(x, y, 
      method = "rf",
      trControl = trainControl(method = "oob"),
      importance = TRUE,
      verbose = TRUE,
      tuneGrid = data.frame(mtry = num.predictors) )


fm <- rf$finalModel


> varImp(f)
rf variable importance

       Overall
var1    100.00
var2    80.14
var3    0.00


> importance(fm)
        %IncMSE IncNodePurity
var2    872.7935      40505276
var1    1021.4707      55682866
var3     273.0168       3078731

Answer 1

As I understood you have only 3 variables. By default varImp function returns scaled results in range 0-100. Var3 has the lowest importance value and its scaled importance is zero. Try to call varImp(rf, scale = FALSE).

Answer 2

Adding to @DrDom's answer, in order to provide further intuition:

The importance scores that varImp(rf, scale = FALSE) gives, is simply calculated by the following: rf$finalModel$importance[,1]/rf$finalModel$importanceSD

This is the feature's mean %IncMSE divided by its standard deviation.