mice package

[mb706]

mice.mids now supports newdata, so could be a pipeop

[mb706]

mice() function: we should have all hyperparameters that are scalar, fixed length, length(ncol) but not length(nrow), since that should always be possible to change. The 'blocks' argument should also allow Selector functions.

Prediction seems to be difficult, see

• Using imputation models already fitted to impute missing values in new individuals? amices/mice#336
• Get at the final model used in the MICE iterations? amices/mice#346

I am not entirely convinced this is all there is, since predictions on the training set give different imputed values than the imputation from training. Is this an RNG thing?