causal-discovery

To learn a causal graph from observational data, use constraint-based or score-based structure learning algorithms like PC, GES, or LiNGAM. These methods infer DAGs or CPDAGs by testing conditional independence relationships within your dataset.

You need a Partial Ancestral Graph (PAG) for causal discovery when latent confounding is plausible. Using the FCI algorithm accounts for hidden common causes that standard PC or GES methods incorrectly assume away under causal sufficiency.

The best way to identify confounders and mediators in cross-sectional data is applying causal structure learning algorithms to infer the causal structure, then analyzing the discovered graph edges to pinpoint likely confounding and mediating pathways.

Yes, you can use causal-learn with pandas and numpy to test a theoretical DAG against empirical measurements. The workflow applies conditional independence tests to validate whether your observed data actually supports the hypothesized causal structure.

Yes, causal structure learning works for non-Gaussian continuous data by using the LiNGAM algorithm. This method estimates a unique Directed Acyclic Graph (DAG) by leveraging non-Gaussianity assumptions instead of relying on standard Gaussian constraints.

The limitation of using PC for social science causal modeling is its strict assumption of no unobserved confounders. FCI overcomes this by generating PAGs, but yields less interpretable graphs with circle edge marks indicating ambiguous causal directions.