Primary Submission Category: Machine Learning and Causal Inference
Causal Inference Post Identification Is Just Functional Estimation: A Critique and Counterexamples
Authors: Abel Mesfin, Elijah Tamarchenko, Katherine Keith, Rohit Bhattacharya,
Presenting Author: Abel Mesfin*
The computation of causal parameters, often expressed in terms of potential outcomes, is a fundamental task in the empirical and social sciences.
Here, we provide a critique of the common adage: Once the identifying functional is obtained for a target causal parameter, causal inference reduces to a pure statistical or functional estimation problem. While this remains true in idealized scenarios, we highlight how common statistical and machine learning practices for dealing with real-world challenges in functional estimation—variable selection for high-dimensional settings and class weighting for handling class imbalance—can, in fact, be counterproductive for the purposes of causal effect estimation. In particular, we show that steps in functional estimation can sometimes increase variance or even introduce bias in causal effect estimates by effectively “undoing” the prior step of causal identification. We ground these arguments in well-known results from the literature on causal graphs and causal discovery, provide empirical evidence of our claims through simulations, and ultimately, synthesize these results into concrete recommendations for handling the bias-variance tradeoff in functional estimation while
being mindful of causal identification.