Closed idg23 closed 1 year ago
The outcome model [nonparametric regression of the response surface $E[Y | X=x]$) is fit on subsets of the population defined by their treatment status, so including the treatment indicator is superfluous (it would be collinear with the intercept).
For details, see Imbens 2004 and subsequent DoubleML papers.
Great package! Very helpful for quick DR computation using ML methods. Thanks. I had one quick question: shouldn't the outcome regression include the treatment indicator along with the the covariates ? I didn't spend enough time on going through the back-end functions but if the treatment indicator is not included in the outcome regression as a feature, aren't we doing something similar to DML then? Many thanks.