Returns a learner that fits ridge regression on the control units and predicts
for all units. Ridge has a closed form (no iterative fitting) and its
shrinkage keeps predictions stable when there are many covariates — the
regime where an unregularized fit overfits and the fit-once shortcut loses
exactness.
Usage
riposte_ridge_learner(lambda = NULL)
Arguments
- lambda
ridge penalty. If NULL, chosen by leave-one-out CV on the
control units each time the learner is called. A fixed lambda (for
instance the CV value from the observed controls) is faster and equally
exact under refit-per-permutation.
Value
a function (X, Y, control) -> Yhat.