A1243
Title: Invariant causal prediction under causal insufficiency
Authors: Marco Borriero - University of Florence (Italy) [presenting]
Monia Lupparelli - University of Florence (Italy)
Giovanni Maria Marchetti - University of Florence (Italy)
Veronica Vinciotti - University of Trento (Italy)
Abstract: Identification of causal relationships among variables, often represented by a directed acyclic graph, is the core of causal discovery. Recently, significant attention has been devoted to methodologies based on the causal invariance principle, which enables the identification of the causal parents of a target variable by exploiting the stability of causal effects across different experimental settings. However, in the more realistic case of causal insufficiency (i.e., when some causal parents are unobserved), the induced graph over the observed variables may no longer be a DAG, and it may not be unique, complicating causal inference. The induced graph and the corresponding structural causal model for relevant configurations of latent parents are derived, and causal invariance is shown to still hold on the induced model for a multivariate Gaussian target. Necessary and sufficient conditions for testing causal invariance are formally determined.
