Estimate ATE of multi-arm treatment on a survival outcome

[einsley1993]

Hi! I would like to ask if the package accommodates the scenario of multi-arm treatment and survival outcomes? I tried out some simulations for a binary treatment and survival outcome. The package worked very well. I'm wondering if it can take multi-arm treatment and survival outcome. There is a multi_arm_causal_forest() function, but it's for non-survival outcomes.

Also, if the dataset is from observational study, there are confounders. In causal_survival_forest(), I specified sample.weights to be the estimated IPTW. For the 5-arm treatment, I estimated propensity score from a multinomial model. But there is also a W.hat option in causal_survival_forest(). Is it supposed to be the PS assuming the treatment is binary?

Thanks for your answers in advance!

[erikcs]

Hi @einsley1993,

Causal survival forest estimates the effect of a binary treatment. If you want to estimate effects for multiple treatment arms, you can do so by repeatedly calling the estimator with one-vs-all encodings of the treatment variable. For example, for treatment k, set W=1 if the k-th arm was assigned and W=0 otherwise. You can use a probability_forest to estimate the propensity scores. This is actually what grf would do before the custom-built multi-arm causal forest.

For your other question, you are right that you can supply the propensity score via W.hat (if this is not provided, grf estimates it). Causal survival forest estimates treatment effects under the same identifying assumptions as causal forest (i.e, unconfoundedness), but with right-censoring due to ignorable drop out added on top.