Stable Blanket with Hidden Variables and Cycles

Hanqing Xiang's new paper, 'Stable Blanket with Hidden Variables and Cycles,' addresses a critical gap in stabilized regression. Traditional stable blanket theory assumes structural causal models (SCMs) without hidden variables or feedback loops—limitations that don't hold in many real-world systems. The paper systematically extends graphical characterizations to three generalized settings: models with hidden variables, models with causal cycles, and models with both. For hidden variables, Xiang uses acyclic directed mixed graphs (ADMGs) and m-separation to define Markov blankets, introducing the novel concept of an 'intervened sub-district' to describe how interventions affect predictor sets. For cycles, the work leverages σ-separation and treats strongly connected components (SCCs) as basic units, working with directed graphs (DGs) and directed mixed graphs (DMGs). The combined framework handles both challenges simultaneously.

Key results include graphical characterizations of stable frontiers and stable blankets, along with conditions for conditional independence of the response from intervention variables given a suitable predictor set. The paper also explores minimality and uniqueness of such sets, offering theoretical guarantees missing in prior work. At 40 pages, this is a thorough contribution to causal inference methodology. Practitioners in econometrics, epidemiology, and AI safety—where confounders and feedback are common—will gain tools to identify intervention-stable predictors even in complex latent-variable, cyclic environments. The work opens the door to more robust domain adaptation and invariant learning in non-idealized settings.