Observed correlations can be explained in multiple ways: variables may directly influence each other, share a hidden common cause, or be subject to selection bias. Each explanation corresponds to a different causal graph, but data alone often cannot tell these graphs apart. A new paper by Ryan Carey and colleagues introduces a graphical object that captures exactly when such causal structures are indistinguishable.

Named the selected-marginalized directed graph (smDG), this structure represents the equivalence class of causal graphs that cannot be distinguished even when researchers are allowed to perform arbitrary interventions on the observed variables. The authors prove that two directed acyclic graphs with latent and selected vertices share the same smDG if and only if they are indistinguishable under such interventions.

The paper also provides a separation criterion for smDGs that is sound and complete for conditional independence under passive observations, serving as an analogue to the familiar d-separation criterion for DAGs. In addition, the authors give a series of sufficient conditions under which two causal structures remain indistinguishable when only passive observations are available. This offers a practical way to know which causal differences are fundamentally beyond experimental reach.