`predict` fails if formula method is used with deleted variables

[adatum]

In a multi-class classification problem, trying to get the class probabilities from a model object trained with the formula method with deleted variables (i.e. -variable) fails with predict when newdata is not specified.

Reproducible example:

fit.lda <- train(Species ~ . -Petal.Length, 
  data = iris, 
  preProcess = c("center", "scale"), 
  trControl = trainControl(method = "repeatedcv", 
    number = 10, 
    repeats = 3, 
    classProbs = TRUE, 
    savePredictions = "final", 
    selectionFunction = "best", 
    summaryFunction = multiClassSummary), 
  method = "lda", 
  metric = "Mean_F1")

then

predict(fit.lda, type = "prob")

fails with:

Error in predict.lda(modelFit, newdata) : wrong number of variables

If the -Petal.Length is omitted in the train formula, there is no error.

This error occurs for some models (lda, knn, svmRadial, mlpML) but not others (nnet, rf, xgbTree).

=====

Workaround

If newdata is specified with the training data, it seems to succeed in retrieving the class probabilities. Is this the correct way of doing it?

In the documentation for predict:

newdata: an optional set of data to predict on. If NULL, then the original training data are used

In the past, I always omitted newdata for obtaining information from the trained model, and included the parameter when using on new testing data.

[adatum]

I'm not sure the workaround is correct. According to http://stats.stackexchange.com/a/112052/130941 predict might behave differently depending on whether newdata is specified or not.

[topepo]

The issue is that an extra column is coming along for the ride (Petal.Length) and some model functions are tolerant to that and others are not.

I'll make a fix for the issue.

[topepo]

This should work find in the devel version and also the one that I just sent to CRAN.