On Theoretical Identifiability of Discrete Latent Causal Graphical Models

This paper considers a challenging problem of identifying a causal graphical model under the presence of latent variables. While various identifiability conditions have been proposed in the literature, they often require multiple pure children per latent variable or restrictions on the latent causal graph. Furthermore, it is common for all observed variables to exhibit the same modality. Consequently, the existing identifiability conditions are often too stringent for complex real-world data. We consider a general nonparametric measurement model with arbitrary observed variable types and binary latent variables, and propose a double triangular graphical condition that guarantees identifiability of the entire causal graphical model. The proposed condition significantly relaxes the popular pure children condition. We also establish necessary conditions for identifiability and provide valuable insights into fundamental limits of identifiability. Simulation studies verify that latent structures satisfying our conditions can be accurately estimated from data.

💡 Research Summary

The paper tackles the fundamental problem of identifying the full causal graphical model when latent variables are present, focusing on discrete (binary) latent variables and allowing observed variables of arbitrary type. Traditional identifiability results for discrete latent models often rely on strong assumptions such as multiple “pure children” per latent variable, restrictive structures on the latent causal graph (e.g., trees or triangle‑free graphs), or the existence of a mixture oracle that reveals the order of marginal modes. These conditions are rarely satisfied in real‑world settings where observed variables share the same modality and the underlying causal structure can be dense.

The authors adopt a non‑parametric measurement model: the observed variables X₁,…,X_J are noisy measurements of the latent binary variables H₁,…,H_K, with no edges among observed variables and no edges from observed to latent variables. The overall edge set is partitioned into Λ (edges among latent variables) and Γ (the bipartite latent‑to‑observed graph). They assume the causal Markov property, faithfulness, and a non‑degeneracy condition that guarantees (a) full support over all 2^K latent configurations, (b) distinct conditional distributions P(X_j | pa(X_j)=h) for different latent parent configurations, and (c) at least one observed child per latent variable.

The central contribution is the “double‑triangular” graphical condition on Γ. Roughly, Γ must contain two distinct triangular sub‑structures: one latent variable connects to two observed variables, and a second latent variable connects to each of those observed variables together with two additional observed variables, forming two overlapping triangles. Under this condition (Theorems 1 and 2), the authors prove that the entire model—latent dimension K, the latent causal DAG Λ (up to Markov equivalence), the bipartite graph Γ (exactly), the latent prior π, and all conditional distributions P(X | H)—is identifiable from the marginal distribution P(X). This condition dramatically relaxes the pure‑children requirement (only two observed children per latent variable are needed) and imposes no structural constraints on Λ, allowing Λ to be any DAG, from a tree to a complete graph, even with isolated latent nodes.

In addition to sufficient conditions, the paper establishes necessary conditions (Theorems 3 and 4). Specifically, each latent variable must have at least three observed children, and the columns of Γ must be non‑nested (no column is a subset of another). While these necessary conditions are weaker than the sufficient double‑triangular condition, they delineate the fundamental limits of identifiability for binary latent models.

The authors validate their theory through extensive simulations. Randomly generated latent graphs Λ and bipartite graphs Γ that satisfy the double‑triangular condition are recovered with near‑perfect accuracy, whereas violations of the condition lead to systematic failures. They also apply the method to a real educational assessment dataset, successfully uncovering a complex latent skill hierarchy that prior methods based on pure children could not identify.

The paper concludes by discussing extensions. The binary latent assumption is crucial for the technical arguments; extending to multi‑valued latent variables would likely require additional constraints. Computational aspects of estimating the model (e.g., EM or variational inference) are left for future work, as are investigations of intermediate graphical patterns that might bridge the gap between the necessary and sufficient conditions.

Overall, this work provides a significant theoretical advance in latent causal discovery: it offers a practically checkable, much weaker graphical condition that guarantees full identifiability of discrete latent causal models without imposing restrictive parametric forms or severe graph‑structural assumptions. The results broaden the applicability of latent causal modeling to a wide range of domains where observed data are homogeneous in modality but underlying causal mechanisms are complex.