Partially Observed Structural Causal Models Introduced
A recent publication on arXiv presents Partially Observed Structural Causal Models (POSCMs), which enhance structural causal models (SCMs) by formalizing causal frameworks where hidden contexts influence both the interaction structure and the mechanisms affecting observed variables. POSCMs facilitate a hierarchy of interventions that includes both node- and edge-level contexts as well as endogenous variable interventions. To support precise edge interventions, the authors utilize a Kolmogorov-Arnold-Sprecher edge-functional decomposition, providing a theorem that represents each node's mechanism as a sum of its parents' univariate functions, leading to a clear parametrization of dyadic functional contributions. The paper also outlines an identifiability theory that specifies which intervention families can effectively separate structure formation from mechanisms. This work is documented as arXiv:2605.03268v1.
Key facts
- POSCMs extend structural causal models (SCMs) to include latent contexts.
- POSCMs allow interventions on nodes, edges, and endogenous variables.
- Kolmogorov-Arnold-Sprecher decomposition is used for edge interventions.
- Identifiability theory clarifies which interventions disentangle structure from mechanisms.
- Empirical validation is performed in a biological context.
- The paper is available on arXiv with ID 2605.03268v1.
Entities
Institutions
- arXiv