Counterfactual Values for Each Treatment in multi_arm_causal_forest

[absbazz-43]

Thank you for creating this excellent package! I am using multi_arm_causal_forest for my project and facing a challenge in obtaining counterfactual values for all treatments.

Currently, the grf package provides individual treatment effect (ITE) estimates by comparing each treatment with the control (reference group). However, I would like to retrieve the counterfactual values for all treatments for each individual, not just the estimates relative to the control group.

Is there a way to extract counterfactual values for all treatments directly using grf? If not, would it be possible to add this feature in future updates?

[erikcs]

Hi @absbazz-43, you can use the $\tau(X_i)$ predictions from multi_arm_causal_forest, along with the propensity score and conditional mean estimates, to back out predictions of the counterfactual arm values $E[Y_i(k) | X_i]$, the docstring for that function has an example of this in the middle: https://grf-labs.github.io/grf/reference/multi_arm_causal_forest.html#examples

grf's sister package policytree also has a utility function, conditional_means that does the same, i.e. you could do policytree::conditional_means(mc.forest) where mc.forest is your fit multi-arm causal forest. This returns a matrix where the k-th colum contains predictions of $E[Y_i(k) | X_i]$.