Title: A causal framework for distribution generalization

Abstract: We consider the problem of predicting a response from a set of covariates when the test distribution may differ from the training distribution. Motivated by the idea that such differences may have causal explanations, we consider a class of test distributions that emerge from interventions in a structural causal model, and focus on minimizing the worst-case risk over this class. Causal regression models, which regress the response variable on all of its direct parents, remain valid under arbitrary interventions on any subset of covariates, but they are not always optimal in the above sense. For example, in linear models, for a set of interventions with bounded strength, alternative solutions have been shown to be minimax prediction optimal. We introduce the formal framework of distribution generalization that allows us to analyze the above problem in partially observed nonlinear models for both direct and indirect interventions on the covariates. It takes into account that, in practice, minimax solutions need to be identified from observational data. Our framework allows us to characterize under which class of interventions the causal function is minimax optimal. We prove several sufficient conditions for distribution generalization and present corresponding impossibility results. We propose a practical method, called NILE, that achieves distribution generalization in a nonlinear instrumental variables setting with linear extrapolation. We prove consistency and present empirical results.