vsyrgkanis

Try setting linear_first_stages=False. If True then we create some extra featurizations to ensure consistency of the estimation and I suspect these extra features crwation is the problem

Try min_var_leaf_on_val=True and min_var_fraction_leaf=0.05 (or sth like that). causal forest involves an inversion of a local matrix defined by the forest neighborhood (unlike forestdrlearner). The default params do not impose...

You have four treatments while the GradientBoostingRegressor that you use for model_t does not support multiple outcomes. You should wrap it with a MultiOutputRegressor.

Nope its currently single T. So you T variable can only he a column or a flat vector. Working on the extension and will be in the next release. For...

@skyetim @hhu1 `CausalForestDML` can now support both multiple treatment and multiple outcome. `LinearDML` has always been supporting multiple treatment. However, if you have multiple binary treatments and you want to...

I think results are reasonable and const marginal cate is the impact of each term in the composite treatment. CIs are wider because you are fitting a much more flexible...

In the final equation there will be a separate theta(X) for each treatment, i.e. theta1(X)*T1 + theta2(X)*T2 + theta12(X)*T12

This is more dependent on developments in dowhy. When we built the wrapper dowhy did not support multiple treatments and hence we prohibited it. If they have since generalized, then...

Its a method we developed in the context of the library. There is no paper on it. There are papers on domain adaptation that use the inverse propensity weighting approach...

Unfortunately we have yet to implement support for sparse matrices in our estimators. It is an important feature.