Introduction
Causal discovery—the task of inferring cause-and-effect relationships from purely observational data—is a cornerstone of modern data science. However, most algorithms rely on the assumption of "causal sufficiency," assuming all measured variables account for every relevant factor. In practice, hidden factors or "unobserved variables" often confound these relationships. This paper, presented at UAI 2021, introduces CAM-UV (Causal Additive Models with Unobserved Variables), a framework designed to identify causal links in nonlinear systems even when the data is incomplete.
The Challenge of Unobserved Variables
Standard functional models often fail when hidden variables are present because they cannot distinguish between direct causation and correlation induced by a hidden common cause. The authors categorize these hidden influences into two types: Unobserved Causal Paths (UCP) and Unobserved Backdoor Paths (UBP).
In nonlinear additive models, these paths leave a distinct "fingerprint" in the residuals of regression. For instance, in a UCP where $X \rightarrow \text{Hidden} \rightarrow Y$, the residual of $Y$ after regressing on $X$ will never be fully independent of $X$. CAM-UV exploits this mathematical limit to detect when a relationship is being "muddied" by unobserved factors.
The CAM-UV Methodology
CAM-UV extends Causal Additive Models (CAM) by incorporating a two-phase estimation process and a conservative marking system.
- Sink Identification: The algorithm identifies "sinks"—variables that do not cause any others in a given subset—by testing for independence between residuals using the Hilbert-Schmidt Independence Criterion (HSIC).
- Two-Phase Estimation: It first identifies candidate parents for each variable and then prunes them to ensure only direct causes remain.
- Honest Uncertainty: When the algorithm detects a connection influenced by a UCP or UBP, it marks the edge as "dashed." This signifies that while a dependency exists, its exact causal nature is ambiguous due to hidden variables.
Experimental Evaluation
The researchers validated CAM-UV using synthetic datasets and simulated fMRI brain activity data.
- Artificial Data: In tests with ten observed variables and multiple hidden confounders, CAM-UV achieved significantly higher precision than traditional methods like FCI or PC.
- fMRI Data: Using the "sim2" dataset, CAM-UV demonstrated remarkable robustness. As more variables were intentionally hidden, its precision remained stable, whereas other methods saw a sharp decline in accuracy.
Conclusion
CAM-UV represents a significant step forward in making causal discovery more reliable in real-world scenarios. By leveraging the unique properties of nonlinear systems and adopting a conservative approach to uncertainty, it avoids the "false positives" that plague simpler models. While computationally intensive, its ability to flag unobserved influences provides a more honest and useful map of causal structures.
Summary updated