TL;DR
This paper introduces LSNM-UV, the first causal discovery framework capable of identifying causal graphs from observational data when both hidden confounders are present and the noise is heteroscedastic (meaning causes modulate both the mean and the variance of their effects). By moving beyond traditional additive noise assumptions, the proposed two-stage algorithm successfully recovers true causal directions where existing state-of-the-art methods fail.
What Problem Does This Paper Address?
Most functional causal discovery methods rely on two major simplifying assumptions: causal sufficiency (assuming there are no unobserved hidden variables confounding the system) and additive noise (assuming a cause only shifts the mean of its effect, leaving the variance constant).
In the real world, both assumptions frequently fail. Hidden variables are ubiquitous, and heteroscedasticity—where the variance of an effect changes depending on the value of its cause—is a common feature of physical, biological, and economic systems. While recent methods have tackled hidden variables under additive noise, or heteroscedastic noise under causal sufficiency, this paper addresses the challenging intersection: discovering causal structures in the presence of both hidden variables and location-scale (heteroscedastic) noise.
Methods
The authors propose a two-level Location-Scale Noise Model with Observed and Hidden Variables (LSNM-UV).
- Noise Reduction Trick: The key mathematical insight is that the location-scale structure allows the effects of hidden variables to be absorbed into an "effective noise" term for each observed variable. This reduces a complex latent-variable problem into a noise-dependence testing problem on the observed margin.
- Residual Characterization: By fitting Generalized Additive Models for Location, Scale, and Shape (GAMLSS) and performing Hilbert-Schmidt Independence Criterion (HSIC) tests, the framework analyzes residual independence patterns to distinguish directed edges, bidirected edges (representing hidden confounders), and non-edges.
- The LSNM-UV Algorithm: A two-stage algorithm. Stage 1 (Base) iteratively identifies "sinks" to extract parent candidates and flags potentially confounded pairs as "invisible." Stage 2 (
checkVisible) re-examines these invisible pairs using a broader search set to resolve true causal directions even when traditional "bow-free" graph assumptions are violated.
Datasets
The method was evaluated entirely on synthetic datasets designed to mimic complex real-world structures:
- Graphs were generated using Erdős–Rényi random directed acyclic graphs (DAGs) with $p=10$ observed variables.
- These were augmented with 2 hidden common causes (creating unobserved backdoor paths) and 2 hidden intermediates (creating unobserved causal paths).
- Data generation utilized highly nonlinear function families (RBF, tanh, softplus, and logarithmic) to define the location and scale mechanisms across sample sizes ranging from $n=200$ to $n=1000$.
Key Insights and Ideas
- Heteroscedasticity as a Feature, Not a Bug: Instead of treating changing variance as a nuisance, LSNM-UV exploits the location-scale noise structure to break the symmetry between cause and effect, enabling causal direction identification where constraint-based methods fail.
- The Power of Effective Noise: Showing that hidden variable effects can be cleanly isolated into an effective noise term means researchers don't need to explicitly model or reconstruct the latent variables to understand their independence structure.
- Moving Beyond Bow-Free Graphs: A "bow" occurs when two variables are connected by both a directed edge and a bidirected edge (a direct causal link plus a hidden confounder). While full graph recovery still requires a bow-free assumption, the authors proved that "visible parents" can act as a side channel to resolve causal directions even within bows.
Key Results
- Superior Directed Edge Recovery: LSNM-UV significantly outperformed additive baselines on heteroscedastic data. It achieved the highest directed edge F1-score across all sample sizes, reaching 0.68 at $n=1000$ (with 0.66 precision and 0.71 recall).
- Failure of Additive Baselines: In contrast, the additive baseline CAM-UV managed a directed F1-score of only 0.13 at $n=1000$, because its additive residuals could not account for parent-dependent variance. Standard constraint-based methods like FCI and BANG remained below 0.04 F1.
- Conservative Confounder Detection: For bidirected edges (identifying hidden confounders), LSNM-UV achieved the highest precision (0.21) but suffered from low recall (0.07), showing a conservative tendency to avoid false positives at the expense of missing some latent links.
Limitations and Future Work
- Low Bidirected Recall: The primary limitation is the low recall for recovering bidirected edges, meaning the algorithm often misses hidden confounders in finite samples.
- Real-World Validation: The current evaluation is restricted to synthetic data. Future work will focus on improving bidirected edge detection and validating the method on real-world observational datasets in fields like biomedicine, education, and economics.
Summary updated