Doubly-robust scores for multi_arm_causal_forest is incorrect

[halflearned]

@erikcs and I found the following error in get_scores.multi_arm_causal_forest.

The expression for the IPW denominator, computed here is only valid for the case of two arms, and incorrect for the case of multiple arms. This gist shows that the error affects the coverage of estimates based on this method.

I ran additional simulations and it seems that the implementation of policytree::double_robust_scores is correct. A correct version of this method should satisfy:

mgrf <- multi_armed_causal_forest(X, Y, W)
pt.scores <- policytree::double_roubust_scores(mgrf)  # AIPW scores for E[Y(w)] for w = 1, ... K
mgrf.scores <- get_scores(mgrf). # AIPW scores for E[Y(w)] - E[Y(1)] for w = 2, ..., K
pt.scores[,-1] - pt.scores[,1] == mgrf.scores 

[erikcs]

This is fixed in #1021. Leaving this issue open until the next CRAN release.