Include moderators that are no confounders

[FlorianNachtigall]

Hi!

Is it possible to include variables that only moderate but do not confound the relationship between T and Y?

Using the CausalForestDML, I would like to include an additional moderator to explain the heterogeneity of the treatment effect, but since it is also a mediator, I can't include it in X. I hope this makes sense.

Any thoughts and workarounds would be greatly appreciated. Thanks in advance!

[kbattocchi]

If you want to use it for heterogeneity you could still add it to X while changing your model_t to something like a pipeline where the first step is a ColumnTransformer that drops that column and the second step is your actual model.

However, if the variable is downstream of T, then such a CATE model will probably not be of much use in a policy setting, for example, because you can't make treatment decisions based on the unknown value of that variable at treatment time.