Replies: 1 comment
-
How did you manage to add variables that affect the target but not the treatment? From all I've seen of EconML models, target and treatment get fed the same sets of features (X and W), only for the effect do you get to choose which features get used (only X). |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hello, I am currently facing a few Issues I would need your help with.
My use case is to build an Uplift Model based on Observational Data. Therefore I am estimating the CATEs (ITEs) with DoWhy/EconML. One step before, I specified the Causal Model with all confounding-variables, that influence treatment & target. Moreover I added further features to the graph that influence the target but not the treatment. Then, I used different estimators like (x-learner, t-learner, drl learner, dml,-learner & causal forest).
After having estimated the cates and having a look on a validation dataset, the CATEs does not find an uplift. I am looking at the bins of the scores of the x-axis and then compare the real mean target per bin seperatet between target & control group (as it can be seen in the screenshot)
After having a closer look on the distribution of the effect modifiers per bin (assuming that the effect variables change the causal estimate, so they should differ in the different ranges of the CATE) I was extremely wondering that the means of each effect modifier per bin is almost the same.
The Example comes from an x-learner but looks similar for a t-learner.
What should I change/tune in order to being able to see an uplift for the different CATE Estimates? More/less variables? Other estimation methods? I really want to see the heterogenous treatment effects show that an one end of the scale, the difference between treatet & control in the target variable is the biggest/lowest
Is it correct that the mean of the effect modifiers does not differ between the bins of the heterogenous CATEs? If yes,why?
And my last question would be if there is existing an example for an uplift model with dowhy/econml on observational data?
Thanks a lot in advance for your help. I am really appreciate your work and I really want to build an successful model.
Beta Was this translation helpful? Give feedback.
All reactions