6.5 - Doubly Robust Methods, Matching, Double Machine Learning, and Causal Trees

[@RayRay-yt5pe]

For real man, thank you so much for everything you have done!!. You have made this stuff really accessible for folks like me!

[@sichenghao3747]

Underrated channel!

[@xyztnull5799]

Thank you very much for the course! I learned a lot here. For double machine learning (DML), I've been reading papers recently. My understanding, which might be incorrect, is that the key point is that stage 1 and stage 2 use two different samples of data. That's, split the whole data into at least 2 samples, then use the first sample of data to estimate Y_hat(.) and T_hat(.) in stage 1; in stage 2, apply the functions Y_hat(.) and T_hat(.) on the second sample of data and then estimate ATE (which is ATE_1). After that, switch sample 1 and 2, and repeat stage 1 and stage 2, you get ATE_2. Finally, average ATE_1 and ATE_2. This data-splitting distinguishes the DML from any other plug-in machine learning estimators. Hope my understanding be correct. Thank you for your excellent course again!

[@BradyNealCausalInference]

That all sounds right. However, you can do sample splitting without estimators as well. For example, you can fit the model for a COM estimator with one sample and then evaluate the estimate with another sample. This kind of sample splitting is also heavily emphasize in the causal trees/forests papers as well. And, as an aside that I'm guessing you already know, it doesn't have to be two samples and then an average between the two; it can be done with k "folds" just like k-fold cross validation.

[@catragmitch]

Hi Brady, Thanks a lot for the video, very informative. I have a question around the nuances between DR and DML - what are the benefits/drawbacks of one over the other? Although I feel I understand the implementation, I am missing any intuition on when to use them, or how their theoretical "guarantees" differ. Thanks in advance!

[@michelspeiser5789]

That causal tree example though 😂

[@paulhowrang]

Wonderful lectures....goldmine

[@RobertWF42]

Thanks Brady! If *both* the conditional outcome *and* propensity score models are inconsistent, how well does the doubly robust model perform?

[@RobertKwapich]

For Double Machine Learning, what happens when we want to estimate the heterogenous treatment effects? The way I understand it, it that we'd have additional variable "X" that impacts (has a directed edge) on Y, and no edge to T. In stage 1 would we fit a model to predict Y from both W and X?