|
Abstract:
|
There is growing interest in estimating heterogeneous treatment effects in experimental and observational studies with application ranging from personalized medicine to online advertisement recommendation systems. The estimate of interest is the Conditional Average Treatment Effect (CATE) function which is defined as the difference between the response under treatment and the response under control. In this talk, we will introduce a new meta-algorithms that can take advantage of any machine learning or regression method, such as BART, Random Forest or KNN, to estimate the CATE. As opposed to other estimation strategies, which often mimic the response functions separately, this meta-algorithm estimates the treatment effect directly. We will show that this method is very adaptive to the underlying data generating process, achieves a minimax optimal rate when combined with KNN, and that it is in particular advantageous when the treatment effect is smooth, or the number of control units is large. We then apply it to Random Forest, and we show how a simple modification of this general meta-algorithm can often be used for estimating valid confidence intervals for both the CATE and individual treatment effects.
|
