NeurIPS 2020

General Control Functions for Causal Effect Estimation from IVs

Meta Review

The reviewers and myself are in agreement that the paper proposes an interesting methodological approach for the instrumental variable regression problem without the additive separability assumption, that is widespread in the literature. The approach also combines interesting ideas from both the econometrics/statistics literature (control functions) and the machine learning literature (autoencoders). The reviewers continue to have concerns on the relation of the identification results to prior literature. The authors tried to address this in the rebuttal and point to a relevant discussion in the paper. However, the short discussion in the paper and in the rebuttal is not sufficient. The authors are therefore strongly encouraged to make the relationship to prior identification theorems in the control function literature and relationship to prior assumptions more concrete, elaborate and technical. Despite this drawback, the paper offers a strong methodological contribution that works well in practice and the theoretical component seems to mostly be a secondary sanity check that the method has theoretical grounding. Given this I am willing to disregard the non-elaborate comparison to the prior identification theorems and recommend that this paper be accepted.