In the ideal world of causal inference, we assume "causal sufficiency"—the belief that we have measured every relevant variable that might influence our system. In reality, this assumption is almost always violated. Unobserved factors, or latent confounders, lurk in the background, creating spurious correlations that look like cause-and-effect relationships.
To navigate this "causal fog," researchers have developed a sophisticated family of graphical models known as Ancestral Graphs. Unlike standard Directed Acyclic Graphs (DAGs), which require every common cause to be explicitly drawn, ancestral graphs allow us to reason about what we can't see by encoding the footprints that latent variables leave behind.
The Limits of Causal Sufficiency
When we use a standard DAG, a correlation between $X$ and $Y$ must be explained by a path: either $X$ causes $Y$, $Y$ causes $X$, or they share a measured common cause $Z$. If a hidden variable $U$ influences both, the DAG is technically incomplete.
As Debo Cheng and Jiuyong Li note in Data-Driven Causal Effect Estimation Based on Graphical Causal Modelling: A Survey, assuming causal sufficiency when it doesn't exist leads to biased estimates. If we don't know $U$ exists, we might mistakenly conclude that $X$ directly causes $Y$.
Representing the Unseen with Mixed Graphs
The breakthrough in this field is the use of Acyclic Directed Mixed Graphs (ADMGs) and Maximal Ancestral Graphs (MAGs). These models introduce new types of edges to represent hidden structures:
- Directed Edges ($X \to Y$): $X$ is a cause of $Y$.
- Bidirected Edges ($X \leftrightarrow Y$): There is an unobserved common cause (a latent confounder) influencing both $X$ and $Y$.
By using bidirected edges, researchers like Ilya Shpitser and Eli Sherman can model complex systems—such as personalized effects in causal pathways—without needing to measure the state of every hidden variable. This abstraction is powerful because it allows for the formal identification of causal effects: determining if a causal question can be answered using only the data we actually have.
From Uncertainty to Partial Ancestral Graphs
Often, the data doesn't give us enough information to decide if a relationship is a direct cause or a confounded one. In these cases, we use Partial Ancestral Graphs (PAGs).
A PAG represents an entire "equivalence class" of possible causal structures. It uses a "circle" notation ($X \circ\!\!\to Y$) to indicate uncertainty about the tail of an edge. As Jean-Philippe Pellet and André Elisseeff demonstrate in Finding Latent Causes in Causal Networks, this allows us to be honest about our ignorance. A PAG tells us which parts of the causal structure are certain and which require more data or interventions to resolve.
How We Find the Right "Control" Variables
One of the most practical applications of this theory is deciding which variables to "control for" in a study. In a simple DAG, we use the "back-door criterion." In ancestral graphs, this is much harder because latent confounders can create "invisible" paths of correlation.
Benito van der Zander, Maciej Liskiewicz, and Johannes Textor provided a definitive solution with their work on Constructing Separators and Adjustment Sets in Ancestral Graphs. They developed algorithms that can efficiently find a valid "adjustment set"—the specific group of variables you must hold constant to see the true causal effect of $X$ on $Y$, even when the graph is full of bidirected edges.
This was further unified by Emilija Perkovic and colleagues in their Generalized Adjustment Criterion (GAC), a single rule that works across DAGs, MAGs, and PAGs alike.
Causal Discovery in the Wild
How do we build these graphs if we don't already know the structure? This is the task of Causal Discovery. Algorithms like Fast Causal Inference (FCI) search through data to find conditional independencies and assemble them into a PAG.
This approach is now being used to solve high-stakes real-world problems:
- Healthcare: Yanke Li and others use PAGs to predict hospital-acquired injuries by distinguishing between direct clinical causes and unmeasured patient background factors.
- Genetics: Gongxu Luo and Biwei Huang apply these frameworks to infer gene regulatory networks, where "selection bias" (only certain cells surviving to be measured) acts as a hidden confounder.
- Fairness: Researchers like Karima Makhlouf use these models to determine if algorithmic fairness metrics are even identifiable when sensitive attributes like race or gender are influenced by unobserved socio-economic factors.
Why It Works: The Intuition of Separation
The core intuition behind all these methods is separation. In a standard graph, if two variables are "d-separated," they are statistically independent. In ancestral graphs, this is extended to m-separation (for MAGs) or mu-separation (for dynamical systems, as explored by Søren Wengel Mogensen).
By looking at which variables remain correlated after we condition on others, we can work backward. If $X$ and $Y$ are correlated, and no amount of conditioning on observed variables makes that correlation go away, an ancestral graph framework allows us to conclude that either $X$ causes $Y$ or there is a bidirected edge ($X \leftrightarrow Y$) representing a "confounder blanket" that we haven't yet pierced. This rigorous accounting for what we don't know is what makes ancestral graphs the gold standard for modern, real-world causal inference.