Abstract
Many real-world decision-making tasks require learning causal relationships between a set of variables. Traditional causal discovery methods, however, require that all variables are observed, which is often not feasible in practical scenarios. Without additional assumptions about the unobserved variables, it is not possible to recover any causal relationships from observational data. Fortunately, in many applied settings, additional structure among the confounders can be expected. In particular, pervasive confounding is commonly encountered and has been utilized for consistent causal estimation in linear causal models. In this paper, we present a provably consistent method to estimate causal relationships in the non-linear, pervasive confounding setting. The core of our procedure relies on the ability to estimate the confounding variation through a simple spectral decomposition of the observed data matrix. We derive a DAG score function based on this insight, prove its consistency in recovering a correct ordering of the DAG, and empirically compare it to previous approaches. We demonstrate improved performance on both simulated and real datasets by explicitly accounting for both confounders and non-linear effects.
Introduction
Causal discovery—the task of learning "what causes what" from data—is notoriously difficult when you can't observe everything. Most traditional algorithms make one of two restrictive assumptions: either all relevant variables are observed (no hidden confounders), or the relationships between variables are strictly linear. In complex systems like biology, neither assumption typically holds. Unobserved transcription factors may influence hundreds of genes (pervasive confounding), and these interactions are rarely linear. Ignoring these factors leads to spurious causal links, while assuming linearity misses true signals. The DeCAMFounder framework bridges this gap by addressing the non-linear, pervasive confounding setting.
Methods
The DeCAMFounder framework introduces a three-part solution to recover causal structures from partially observed, non-linear data.
1. The Structural Model
The authors extend Causal Additive Models (CAMs) to include hidden variables ($H$). Each observed variable is modeled as a sum of non-linear functions of its parents, a non-linear function of the hidden confounders, and random noise. This allows for a more realistic representation of complex systems where unobserved factors exert broad influence.
2. Principal Confounding Sufficient Statistics (PCSS)
A key insight from high-dimensional statistics is that if a hidden variable affects many observed variables (pervasive confounding), its influence is concentrated in the top eigenvalues of the data's covariance matrix. DeCAMFounder leverages Principal Component Analysis (PCA) on observed data to estimate these confounding signals without direct observation of the hidden variables. This "blessing of pervasiveness" makes spectral methods highly effective for deconfounding.
3. The DeCAMFounder Score
The framework utilizes a Bayesian score function based on Gaussian Processes (GPs). This score measures the fit of a candidate Directed Acyclic Graph (DAG) after "subtracting out" the confounding variation estimated via PCA. Maximizing this score is proven to asymptotically recover the correct topological ordering of the causal graph.
Experimental Results
The method was evaluated on both synthetic and biological datasets to test its robustness against non-linearity and confounding.
- Simulated Data: Using Erdős-Rényi random graphs ($p \le 1,000$), the researchers tested non-linear trends like sine waves and quadratic functions. DeCAMFounder significantly outperformed linear benchmarks (e.g., LRPC+BIC) and standard non-linear methods that ignore confounders, showing a marked reduction in false positive and false negative edges.
- Ovarian Cancer RNA-seq Data: In a real-world test involving 486 genes, the authors treated 15 known transcription factors as hidden confounders. DeCAMFounder achieved a positive likelihood ratio of 0.72, nearly doubling the performance of standard CAMs (0.42) and tripling that of Vanilla-BIC (0.26).
Conclusion and Future Work
DeCAMFounder demonstrates that spectral deconfounding is robust and relatively insensitive to the exact number of principal components, provided the confounding signal is strong. However, the framework currently does not account for selection bias (M-bias), and the computational cost of Gaussian Processes ($O(p^2N^2 + pN^3)$) may require sparse GP approximations for larger scales. Future research will focus on these scaling challenges and the impact of non-pervasive confounders.
Summary updated