partial ancestral graph (PAG)
Definition
Partial ancestral graph (PAG) is a graphical representation used in constraint-based causal discovery to encode a set of causal structures that are statistically equivalent given the observed data, including cases where unobserved confounding variables may be present. Rather than committing to a single causal graph, a PAG represents the features shared across all graphs in a Markov equivalence class, using a specific set of edge marks to indicate which causal features are identifiable and which remain uncertain. In the context of psychological research, PAGs arise as output from FCI-variant algorithms, which are designed to handle both cyclic causal structures and latent confounders, making them relevant for modeling phenomena such as psychopathology where feedback loops between symptoms are theoretically expected.
Sources: Park et al. (2024)
Related Terms
- constraint-based (1 shared article)
- cyclic causal discovery (1 shared article)
- directed cyclic graph (DCG) (1 shared article)
- statistical network model (1 shared article)
Applications
Partial Ancestral Graph (PAG) and Cyclic Causal Discovery
PAGs are produced by FCI-variant algorithms, which belong to a class of constraint-based cyclic causal discovery methods evaluated for use in psychological research contexts. Simulation evidence indicates that FCI-variant methods, and by extension the PAGs they output, perform less well than autoregressive-based methods when recovering cyclic causal structures from psychologically plausible data.
Sources: Park et al. (2024)
Partial Ancestral Graph (PAG) and Latent Confounding Variables
A core motivation for using PAGs is their capacity to represent uncertainty introduced by unobserved confounders, which cannot be ruled out in typical observational psychological datasets. The FCI algorithms that generate PAGs are explicitly designed to remain agnostic about causal directions and adjacencies that cannot be distinguished given the possibility of latent variables, a condition that becomes especially consequential when sample size and network density vary across simulation conditions.
Sources: Park et al. (2024)
