Confidence intervals for CLATE in instrumental forest?

[lepromatous]

I can't seem to figure out a CI for CLATE in an instrumental forest is computed appropriately - any thoughts or direction-to-be-pointed-in is much appreciated. Thanks so much!

[swager]

Hi @lepromatous, I'm not sure I quite understand the question. Are you trying to get CIs for the average CLATE (like in the average_treatment_effect method), or do you want CIs for \hat{tau}(x) as provided by an instrumental forest? A code snippet outlining what you're aiming to do would be helpful.

[lepromatous]

Thanks for the response! Im trying to get both actually! Project is aimed at showing subgroups which may benefit more/less from a specific treatment (observational data, hence the instrument).

Currently this is in 'proof-of-concept' phase. If I can make sense of it all I will move on to actually tuning this appropriately (parameters here are inappropriately gleaned from internal tune function for causal forest - is there a tune instrumental forest function?)

model -- outcome is binary, instrument is quantiles of a continuous value and meets criteria for ####### strong instrument

test = instrumental_forest(X=iv[,-c(69, 125, 127, 139)], Y=iv[,128], W=iv[,69], Z=as.numeric(as.character(adi.cut)),
sample.fraction=0.5, min.node.size=2,
mtry=24, alpha=0.15221056, num.trees = 6000)

My questions are twofold:

1. what is the CLATE function for instrumental forest? average_treatment_effect(test) errors that it is only implemented for causal_forest.

Can you just predict for OOB CLATE and average everything? a la (seems unlikely to be valid...):
out<-predict(test)[,1]
iv$pred<-out
iv$pred
mean(iv$pred)

If so, to get HTE, just do the same on a subset post predict(test)?

2. How to calculate CIs for CLATE and \hat{tau}(x) while attempting to maintain a 0/1 boundary for the binary outcome?

Thanks so much!!!

[foo-bar-baz-qux]

Also very interested in this, would be great to have average_treatment_effect() work for instrumental forests as well.

[swager]

Apologies @lepromatous for dropping this thread.

1. If there's regular demand for this functionality, we should probably add an averagage_clate function to the package that takes an IV forest as an argument. See Estimate of LATE using instrumental_forest #358 for a discussion of how to do this "by hand". Contributions welcome!
2. One way to make CIs shorter is to use a larger minimum node size to regularize the forest more heavily. If the CIs are outside the [0, 1] interval even after that, then this may be indicative of a weak IV. For issues like these, I'd recommend doing simulation studies with data-generating distributions that look a lot like your real dataset so you can get a better feel for what the forest is doing.

[jtibshirani]

I've marked this as a feature request. Note that this was previously requested in #358.

[kingsharaman]

+1 for this feature! Just run into this thread while trying to solve a similar problem.

[hliebert]

I'd be very interested in this feature as well.