📊 statistics

Causal Additive Models with Unobserved Causal Paths and Backdoor Paths

This paper establishes sufficient conditions for identifying causal directions in causal additive models with unobserved backdoor and causal paths by characterizing regression sets and residuals, and introduces a sound and complete search algorithm that demonstrates competitive performance against state-of-the-art methods.

Original authors: Thong Pham, Takashi Nicholas Maeda, Shohei Shimizu

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thong Pham, Takashi Nicholas Maeda, Shohei Shimizu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out who influenced whom in a group of people. You have a list of suspects (observable variables), but you know there are some secret agents (hidden variables) pulling strings behind the scenes that you can't see. Your goal is to draw a map showing who is the boss and who is the employee.

This paper introduces a new detective tool called CAM-UV-X to solve this mystery, specifically when the "boss" and "employee" relationship is clouded by these secret agents.

Here is the breakdown of the problem and the solution, using simple analogies:

The Problem: The "Invisible Fog"

In the world of data science, there are existing methods (like the old CAM-UV tool) that are good at spotting clear relationships. If Person A clearly influences Person B, the tool says, "Yes, A is the boss of B."

However, these tools hit a wall when there are hidden paths:

The "Bow" Shape: Imagine Person A and Person B are both influenced by a secret agent, Person U. This creates a "bow" shape (A ← U → B). To the old tools, A and B look like they are just friends or unrelated because the secret agent U is messing up the signal. The old tools would say, "I can't tell who is the boss here; it's too foggy."
The "Backdoor" Path: Imagine A influences B, but there's also a secret path where a hidden agent influences both. This creates a "backdoor" that confuses the detective.

The paper argues that the old tools were too pessimistic. They gave up too easily, declaring many relationships "unidentifiable" (unsolvable) when they actually could be solved with a smarter approach.

The Solution: A New Detective Strategy (CAM-UV-X)

The authors, Thong Pham, Takashi Nicholas Maeda, and Shohei Shimizu, built a new algorithm called CAM-UV-X. Think of this as upgrading the detective's toolkit with two new superpowers:

1. The "Residual Scrub" (Regression Analysis)

Imagine you are trying to hear a whisper (the direct influence of A on B) in a noisy room. The noise is the influence of other people.

Old Method: The detective tries to guess the noise and subtract it. If they guess wrong, they can't hear the whisper.
New Method: The paper introduces a way to mathematically "scrub" the data. They look at the leftover bits (residuals) after accounting for other factors.
The Magic: They discovered that even if there is a hidden agent (the "bow"), if you scrub the data in a very specific way, the leftover bits will still show a connection only if there is a direct boss-employee relationship. It's like realizing that even though two people are wearing the same uniform (influenced by the same secret agent), their footprints (residuals) still show who is leading the other.

2. The "Hybrid Detective" (Combining Clues)

Sometimes the "Residual Scrub" isn't enough. The new tool combines two types of clues:

Clue Type A: The leftover bits from the scrub (as described above).
Clue Type B: Standard "conditional independence" (checking if two people are unrelated when you know about a third person).

The Analogy:
Imagine you are trying to prove that Alice is the boss of Bob, but they both have a secret friend, Charlie, who influences them both.

Old Tool: "Alice and Bob are both friends with Charlie. I can't tell if Alice bosses Bob. Case closed."
New Tool (CAM-UV-X): "Wait. I know Alice bosses Bob because if I remove Charlie's influence, Alice and Bob still have a unique connection. Also, I know that Dave (another person) is influenced by Alice but has no connection to Bob. This confirms Alice is the boss."

By mixing these two types of logic, the new tool can solve cases that were previously considered impossible, including the tricky "bow" shapes.

What the Paper Actually Claims

The authors make three specific claims based on their math and experiments:

It's Solvable: They proved mathematically that in many cases where hidden variables exist (specifically "unobserved backdoor paths" and "unobserved causal paths"), you can still figure out who is the boss and who is the employee. You don't need to know the hidden variables to solve the puzzle; you just need to look at the data differently.
The Algorithm Works: They created CAM-UV-X. They proved that this algorithm is sound (it never lies; if it says A is the boss of B, then A really is) and complete (it finds every relationship it is theoretically capable of finding).
It Performs Well: They tested it on fake data (simulated graphs) and real-world sociological data (about fathers' jobs and sons' incomes).
- On the fake data, it was better at finding the correct connections than the old CAM-UV tool.
- On the real data, it corrected mistakes the old tool made, such as removing fake "bidirectional" connections (where the tool thought two people influenced each other equally) and replacing them with the correct one-way direction.

What It Does Not Claim

It does not claim to work for every single possible graph structure. There are still some extremely complex "bows" that might remain unsolvable, but the authors haven't found an example of one yet.
It does not claim to be a medical cure or a financial trading strategy. It is strictly a method for discovering causal structures from data.
It does not claim to replace all other methods, but rather to improve upon the specific "Causal Additive Model" framework used in previous research.

Summary

The paper is like a detective saying: "We used to think that if a secret agent was involved, we couldn't solve the case. But we found a new way to look at the evidence (scrubbing the data and mixing clues) that allows us to solve many of those 'impossible' cases. We built a new tool (CAM-UV-X) that does this, and it works better than the old tools."

Technical Summary: Causal Additive Models with Unobserved Causal Paths and Backdoor Paths

Problem Statement
Discovering causal structures from observational data is a fundamental challenge, particularly when unobserved (hidden) variables are present. Causal Additive Models (CAMs) offer a tractable framework for modeling nonlinear dependencies with additive noise. However, existing theories for causal discovery in CAMs with hidden variables, such as the CAM-UV algorithm (Maeda and Shimizu, 2021), face significant limitations. Specifically, when unobserved backdoor paths (UBPs) or unobserved causal paths (UCPs) exist between two variables, the parent–child relationship is often deemed unidentifiable. This includes "bow" configurations (where two observed variables share a common hidden parent), which are notoriously difficult for causal discovery because hidden confounding obscures direct causal links. Previous approaches, including the Fast Causal Inference (FCI) framework, generally treat these relationships as unidentifiable without additional strong assumptions on the graph structure.

Methodology
The paper proposes a new theoretical framework and a corresponding search algorithm, CAM-UV-X, to address these identifiability gaps. The methodology relies on a hybrid approach combining:

Regression Residual Independence: Characterizing the sets of variables used in Generalized Additive Model (GAM) regressions to determine independence among residuals.
Conditional Independence: Utilizing conditional independence among the original observed variables.

Key Theoretical Innovations:

Refined Visibility/Invisibility Characterization: The authors extend the concepts of "visible" and "invisible" pairs to subsets of observed variables ( $X' \subseteq X$ ). They establish that a pair $(x_i, x_j)$ may be invisible with respect to the full set $X$ (due to UBPs/UCPs) but become visible when a specific variable $x_k$ is removed from the conditioning set.
New Identifiability Conditions:
- Lemma 4 & 5: These lemmas provide sufficient conditions to identify that a variable $x_k$ is an ancestor of $x_i$ or $x_j$ even when the pair $(x_i, x_j)$ is invisible. This is achieved by analyzing whether residuals become independent only when specific variables are included in the regression sets.
- Hybrid Approach (Corollary 6.1 & 6.2): The paper introduces conditions that combine regression residual independence with conditional independence ( $x_k \perp\!\!\perp x_j | x_i$ ). This allows for the identification of causal directions in invisible pairs (e.g., determining that $x_i$ is an ancestor of $x_j$ ) and resolving parent–child relationships in bow patterns, cases where neither regression alone nor standard conditional independence methods (like FCI) succeed.
Algorithm CAM-UV-X: The proposed algorithm builds upon the original CAM-UV. It first runs CAM-UV to obtain an initial estimate, then iteratively applies the new lemmas and corollaries to:
- Correct misclassifications of visible edges/non-edges as invisible (addressing incompleteness in CAM-UV).
- Orient invisible pairs by checking specific regression and conditional independence criteria.
- The algorithm is proven to be sound and complete for identifying visible edges, visible non-edges, and invisible pairs under the Causal Faithfulness Condition (CFC) and a residual-faithfulness assumption.

Results
The paper presents empirical evaluations demonstrating the performance of CAM-UV-X:

Illustrative Examples: In synthetic graphs (Figs. 1a and 1b) designed to expose the limitations of CAM-UV (where CAM-UV incorrectly labels visible edges as invisible), CAM-UV-X successfully recovers the correct causal structures.
Random Graph Experiments: On Barabási–Albert (BA) and Erdős–Rényi (ER) graphs with hidden confounders, CAM-UV-X achieves higher precision and F1 scores in identifying the adjacency matrix and higher recall in identifying ancestral relationships compared to CAM-UV and a baseline adaptation of Schultheiss and Bülmann (2024).
Real-World Data: Applied to a sociological dataset (father's occupation, son's income, etc.), CAM-UV-X pruned spurious bidirected edges found by CAM-UV and correctly identified directed edges consistent with ground truth domain knowledge. It also provided additional information (non-adjacency assertions) for invisible pairs that CAM-UV could not resolve.

Significance and Claims
The authors claim that this work establishes sufficient conditions under which causal directions and parent–child relationships can be identified in CAMs even in the presence of unobserved backdoor or causal paths, a scenario previously considered unidentifiable without extra structural assumptions.

Extension of Prior Work: The results extend the work of Maeda and Shimizu (2021) by identifying parent–child relationships in "bow" configurations and extend Schultheiss and Bülmann (2024) by identifying causal directions beyond their well-specified sets.
Practical Impact: By providing a sound and complete algorithm for these specific cases, the paper aims to improve the practical adoption of causal discovery in settings where hidden variables are prevalent.
Limitations and Future Work: The authors modestly note that their current theory does not fully utilize all constraints implied by d-separation among observed variables. They identify the gap between sufficient and necessary conditions in CAMs with hidden variables as an open problem and suggest extending these hybrid approaches to other classes of causal models as a direction for future research. The paper does not claim to solve all identifiability issues in the presence of hidden variables, but rather to resolve specific, previously intractable cases.