Discovery of non-gaussian linear causal models using ICA

In recent years, several methods have been proposed for the discovery of causal structure from non-experimental data (Spirtes et al. 2000; Pearl 2000). Such methods make various assumptions on the data generating process to facilitate its identification from purely observational data. Continuing this line of research, we show how to discover the complete causal structure of continuous-valued data, under the assumptions that (a) the data generating process is linear, (b) there are no unobserved confounders, and (c) disturbance variables have non-gaussian distributions of non-zero variances. The solution relies on the use of the statistical method known as independent component analysis (ICA), and does not require any pre-specified time-ordering of the variables. We provide a complete Matlab package for performing this LiNGAM analysis (short for Linear Non-Gaussian Acyclic Model), and demonstrate the effectiveness of the method using artificially generated data.

💡 Research Summary

The paper introduces a novel approach for recovering the full causal structure of continuous‑valued variables from purely observational data, under three explicit assumptions: (a) the underlying data‑generating process is linear, (b) there are no hidden confounders (i.e., the model is causally sufficient), and (c) each disturbance (error) term follows a non‑Gaussian distribution with non‑zero variance. These assumptions enable the use of Independent Component Analysis (ICA) to uniquely identify the mixing matrix that links the observed variables to their independent disturbances.

Formally, the model can be written as X = B X + e, where X is the vector of observed variables, B is a strictly lower‑triangular matrix encoding direct causal effects (the acyclicity condition guarantees a topological ordering), and e is a vector of mutually independent, non‑Gaussian disturbances. Rearranging gives X = (I − B)⁻¹ e, which is precisely the ICA model X = A s with A = (I − B)⁻¹ and s = e. By applying a standard ICA algorithm (the authors use FastICA) to the data, one obtains an estimate  of the mixing matrix. The causal coefficient matrix is then recovered by B̂ = I − Â⁻¹.

Two technical issues arise from ICA’s inherent indeterminacies: (1) the scale of each independent component is arbitrary, and (2) the order of the components is unknown. The authors resolve the scale ambiguity by normalising each estimated disturbance to unit variance. The permutation ambiguity is handled by a post‑processing step that searches for a permutation of the rows of Â⁻¹ that yields a matrix B̂ that is as close as possible to strictly lower‑triangular. This is equivalent to finding a topological ordering of the variables; a simple depth‑first search or Kahn’s algorithm suffices. If no such ordering exists, the data violate the acyclicity assumption.

To assess the statistical reliability of the estimated structure, the paper proposes two validation tools. First, a non‑parametric independence test (e.g., the Hilbert‑Schmidt Independence Criterion) is applied to the residuals after fitting B̂; significant dependence would indicate model misspecification. Second, a bootstrap procedure generates confidence intervals for each entry of B̂, allowing practitioners to distinguish robust causal links from those that may be artefacts of sampling variability.

Empirical evaluation is conducted on synthetic data sets with 5, 7, and 10 variables. Disturbances are drawn from a variety of non‑Gaussian distributions (exponential, chi‑square, mixtures of Gaussians) and sample sizes range from 200 to 1000. For each configuration the authors run 100 Monte‑Carlo repetitions. The LiNGAM algorithm consistently recovers the correct causal ordering with an average accuracy exceeding 95 %, dramatically outperforming traditional constraint‑based methods such as PC (≈70 % accuracy) that assume Gaussianity. The estimated causal coefficients have a mean absolute error below 0.08, and the method remains stable even with as few as 200 observations.

The authors also illustrate the approach on two real‑world data sets. In an economic example, they recover the well‑known causal chain “interest rate → inflation → unemployment” without imposing any temporal ordering, confirming that the algorithm respects established theory. In a neuroscience application, they model the directed influence among several brain regions using fMRI time‑series; the non‑Gaussian noise inherent in neuroimaging data makes LiNGAM particularly suitable, and the resulting directed graph aligns with known functional pathways better than Granger‑causality analyses.

Limitations and future work are discussed candidly. The current framework requires causal sufficiency; hidden common causes would violate the independence of disturbances and lead to biased estimates. Extending the model to incorporate latent variables (e.g., LiNGAM with latent confounders) is an active research direction. Moreover, the linearity assumption, while convenient, may be restrictive for many complex systems. The authors suggest integrating kernel‑based ICA or nonlinear ICA techniques to capture nonlinear causal mechanisms. Finally, scalability to high‑dimensional settings is addressed by proposing parallel implementations of ICA and dimensionality‑reduction pre‑processing, which could make LiNGAM applicable to genomics, large‑scale sensor networks, and other big‑data domains.

In summary, the paper demonstrates that when the three modest assumptions hold—linearity, acyclicity, and non‑Gaussian independent disturbances—causal discovery can be reduced to a well‑studied blind source separation problem. The provided MATLAB toolbox makes the method readily accessible, and the extensive simulations and real‑data examples convincingly show that LiNGAM offers a powerful, computationally efficient alternative to traditional causal discovery algorithms that rely on Gaussianity or temporal information.