Out-of-fold predictions exist only for the training rows (and are stored
on the fitted object). For genuinely new data, predict.meep() refits
each base learner on the full training sample – using the frozen
hyperparameters when tune = "once" – then combines the refits with the
stored ensemble weights. Full-data refits are cached lazily in an
environment field on the object and are not serialized.
Usage
# S3 method for class 'meep'
predict(
object,
newdata,
nuisance = c("outcome", "treatment", "mu0", "mu1"),
...
)
Arguments
- object
A fitted "meep" object.
- newdata
A matrix or data.frame of new covariates.
- nuisance
Which nuisance to predict: "outcome", "treatment",
"mu0", or "mu1".
- ...
Unused.
Value
A numeric vector of length nrow(newdata).
Details
Predictions on new data can extrapolate; the cross-fitting honesty
guarantee applies only to the training-row OOF predictions.