Performance of First Stage Models

How does the predictive performance of the first stage models affect confidence in the final stage model?

For context, I am using the ForestDMLCateEstimator with a random forest regressor and random forest classifier for model y and model t, respectively, and I use a grid search to select hyperparameters for my first stage models. The predictive performance of my model y (regressing the outcome on the features aside from treatment) is low (r-squared is in the vicinity of 0.3) but the performance of model t is fairly higher (classification accuracy is ~0.85). Does the poor performance of my model y negatively affect the validity of my treatment effect results? Is there a typical benchmark for performance of the first stage models?

Thank you in advance and thanks for the great documentation!

Start with the ForestDMLCateEstimator documentation and its guidance on first-stage models; review how model y and model t are evaluated in the reported setup. Done means documenting whether the stated R-squared and classification accuracy affect treatment-effect validity and whether a typical performance benchmark exists.