Closed sami-ka closed 9 months ago
I don't think I'd say the assumptions on the DGP are less restrictive for DeepIV, they're just different. The functional form of the relationship between Y and T can be arbitrary, but identifiability depends in a somewhat complex way on the distribution of T given X, W, and Z. For example, since we're trying find g minimizing (Y-∫g(T,X,W)dF(T|X,W,Z))², if the distribution of T is always symmetrical given X, W, and Z, then we can't identify the odd part of g, only the even part (since adding any odd function of T to g will give the same result for the integral).
Hi, I would like to make sure that what I understood of the difference between OrthoIV and DeepIV is correct/complete.
Can we say that : DeepIV makes less restrictive assumptions on the DGP than OrthoIV. It is then more generic in that sense. However it involves the training of neural networks, which may be slower to converge to a good state compared to classic ML models. So : if we decide that the assumptions of the DGP behind OrthoIV is not too restrictive for our task, go with OrthoIV (easier to train) else go with DeepIV. ?
Feel free to correct me or complete the statement.