In the search for "what causes what," the most important factors are often the ones we cannot measure directly. In psychology, we measure test scores to infer "intelligence"; in genetics, we observe gene expressions to infer underlying "regulatory pathways." These unobserved factors are latent variables.
Traditional causal discovery—the process of mapping cause-and-effect from raw data—often treats latent variables as a nuisance or assumes they follow extremely simple patterns. However, recent research has shifted toward identifying complex, hierarchical structures where hidden variables can cause each other, overlap in their influences, and even be influenced by the variables we can see.
The Problem with "Pure" Models
For decades, the standard approach to latent variables relied on the pure measurement model. This assumes that every observed variable (an indicator) is the effect of exactly one hidden cause. If you were measuring "Anxiety," every question on your survey would be assumed to stem only from that single latent trait.
As Ricardo Silva and Ricardo Bezerra de Andrade e Silva point out in Measuring Latent Causal Structure, real-world data is rarely "pure." A single survey answer might be influenced by both "Anxiety" and "Depression." When indicators are "impure"—influenced by multiple overlapping latent factors—traditional factor analysis fails to recover the true causal map.
Furthermore, many frameworks assume that latent variables are "root causes" (nodes with no parents). Alex Markham and Xudong Sun argue in Addressing pitfalls in implicit unobserved confounding synthesis that this ignores biological and social realities where observed actions can change hidden states.
The Geometry of Hidden Causes: Rank Constraints
How do you find a variable you can't see? The key intuition lies in the rank of the covariance matrix of your observed data.
Imagine two observed variables, $X_1$ and $X_2$. If they are correlated, it might be because $X_1 \to X_2$. But if they are correlated only because they share a single hidden cause $L$, the information flowing between them must pass through a "bottleneck." Mathematically, this bottleneck manifests as a rank constraint.
Biwei Huang, Feng Xie, and their colleagues have pioneered the use of these constraints to move beyond simple trees. In Latent Hierarchical Causal Structure Discovery with Rank Constraints, they show that by checking the rank of different subsets of the covariance matrix, we can determine not just that a latent variable exists, but exactly where it sits in a hierarchy.
This approach, further refined in A Versatile Causal Discovery Framework to Allow Causally-Related Hidden Variables by Xinshuai Dong and Yujia Zheng, allows for "RLCD" (Rank-based Latent Causal Discovery). This framework can handle scenarios where observed variables cause each other directly while simultaneously being influenced by a web of hidden factors.
Exploiting Non-Gaussianity and Asymmetry
While rank constraints help find the existence of links, they don't always tell us the direction (e.g., does $L_1 \to L_2$ or $L_2 \to L_1$?). To solve this, researchers look at the "shape" of the noise in the data.
In linear models with Gaussian (bell-curve) noise, the direction of causality is often mathematically invisible. However, if the noise is non-Gaussian, the distribution becomes asymmetric. Feng Xie and Biwei Huang leverage this in Identification of Linear Non-Gaussian Latent Hierarchical Structure. By analyzing these statistical signatures, their LaHME algorithm can orient edges between hidden variables even when those variables have no direct observed children.
This logic extends to even more complex systems. Haoyue Dai and Kun Zhang recently demonstrated in Distributional Equivalence in Linear Non-Gaussian Latent-Variable Cyclic Causal Models that these non-Gaussian signatures can even reveal causal cycles (feedback loops) involving hidden variables, a feat previously thought impossible from observational data alone.
Beyond Trees to Causal Webs
The ultimate goal of this research direction is to move from "latent trees" to "latent DAGs" (Directed Acyclic Graphs). In a tree, there is only one path between any two nodes. In a DAG, multiple paths can exist, representing the messy reality of systems like the human brain or global economies.
Recent work has broken the "one-parent" rule:
- Impure Indicators: Ruichu Cai and Zhifeng Hao use tensor rank conditions in Learning Discrete Latent Variable Structures with Tensor Rank Conditions to model discrete data (like survey results) where one answer is driven by multiple hidden factors.
- Nonlinear Relationships: Lingjing Kong and Eric P. Xing show in Identification of Nonlinear Latent Hierarchical Models that we can recover these hierarchies even when the relationships between variables are complex and nonlinear.
By relaxing the restrictive assumptions of the past, this field is turning causal discovery into a tool capable of mapping the "invisible" architecture of the real world. Readers no longer have to assume a simple structure to get an answer; the data's own geometric and statistical properties now reveal the hidden hierarchy for them.