We completely agree that multi-decision confounding is also important, but

We'll address them all in the paper & answer the key questions here. We will add this discussion in the final version. D and relation to Z & B 2019 (R1) We briefly discussed this in Lines 76-80 and will highlight it further. We can minimize the loss using any model class (e.g. We only use the linearity assumption in our proof to guarantee statistical consistency.