Handle numeric NA values

[hneth]

This PR improves the handling of NA values in numeric predictors (by adding the option of replacing NA by means) and revises corresponding tests.

[pa-nathaniel]

Hi Hans! Just lurking here without fully reading the code and only the description. But is this PR really enabling missing value imputation logic for within FFTrees? If so is it enabled by default or not? For both training and test data or just one?

Sorry if I missed the description in NEWS or readme, but I didn't see it, thanks!

[hneth]

Hi Nathaniel,

good points, of course. Some clarifications to answer your questions:

But is this PR really enabling missing value imputation logic for within FFTrees?

Yes, it’s replacing NA values in numeric predictors by their mean (per predictor), as commonly done in simulation studies. (Other replacement policies could easily be implemented in the same way.)

If so is it enabled by default or not?

Yes, but there are several global constants that allow enabling / disabling functionality for handling NA values in data. The relevant ones here are:

- allow_NA_pred (to dis/allow NA values in predictors)

• replace_NA_num_pred (to dis/en-able replacing NA values in numeric predictors)

Both are currently set to TRUE by default, but generate a bunch of warnings when actually encountering and replacing NA values.

If allow_NA_pred = TRUE, but replace_NA_num_pred = FALSE, NA values in numeric predictors are ignored in FFT-generation (but issue corresponding messages) and classified according to the current choice from fin_NA_options (when reaching a final node). For the competing models, cases with NA values are removed in this case (to avoid crashing).

For both training and test data or just one?

Both training and test data are handled in the same way — and detecting and replacing NA values issues corresponding warnings for both.

Overall, and especially in combination with the default handling of NA values in categorical predictors (as distinct categories), these changes should provide pretty comprehensive options for a variety of cases.

I’d be happy to discuss which of the options should become defaults and which should become user-controlled parameters (of the main FFTrees() function). Also, the new functionality still needs to be tested and documented properly before being ready for release. (If we wanted to be conservative, setting the global NA-related options to FALSE removes all of this immediately and restores the former functionality.)

Looking forward to hear your thoughts,
Hans