Hyperparameter configurations for multiple treatment variables

[mateusnmsouza]

I have a question related to DoubleML applications with multiple treatment/policy variables: https://docs.doubleml.org/stable/examples/py_double_ml_plm_irm_hetfx.html

It seems that the package currently only allows the user to provide a single hyperparameter configuration, even though there are multiple treatment variables. However, in practice, the hyperparameters for one treatment variable might not be the best for the other. In some cases, one could even consider using different algorithms for each treatment variable. Am I missing something? Is there a way of supplying different hyperparameters for each variable?
To be clear, this is not a question about tuning. I have already performed tuning separately for each variable through Optuna.

Thanks for any insights on this issue.

[OliverSchacht]

Hi @mateusnmsouza ,

thanks for raising this interesting question.
When you hand over a learner to the PLR model, indeed the same hyperparameters are used for all treatment columns.
However, you can use the set_ml_nuisance_params() to set individual hyperparameter configurations per learner and per treatment. This function is available for all model classes, see also the Hyperparameter Tuning Section in the docs.

if you want to go beyond this and, say, use different ML models for different treatment propensity predictions, then this is currently only possible when providing the cross-fitted predictions as external predictions.

I hope this helps!

Oliver

[mateusnmsouza]

Yes, this helps. Thanks!