Causal Discovery Algorithms (PC, FCI, LiNGAM)
Also known as: PC algorithm, FCI algorithm, LiNGAM, causal structure learning, DAG learning, Nedensel Keşif Algoritmaları (PC, FCI, LiNGAM)
Causal discovery is a family of algorithms that automatically learn a directed acyclic graph (DAG) describing causal structure directly from observational data. The constraint-based PC and FCI algorithms were developed by Spirtes, Glymour and Scheines (2000), while the LiNGAM model of Shimizu et al. (2006) exploits linear non-Gaussian structure to orient edges.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use causal discovery when you want to explore or hypothesise the causal structure among several continuous or binary variables from observational data, with a reasonably large sample (at least about 100 observations). PC and FCI rely on the Markov condition and faithfulness; FCI is the right choice when latent confounders are plausible. LiNGAM requires a linear model with non-Gaussian, acyclic noise. With many variables the size of candidate parent sets must be bounded, and any discovered graph should be interpreted with domain expertise rather than taken as established causation.
Strengths & limitations
- Learns causal structure directly from observational data when controlled experiments are infeasible.
- Offers complementary engines: PC/FCI for constraint-based testing and LiNGAM for fully oriented graphs under non-Gaussian noise.
- FCI explicitly allows for hidden (latent) confounders rather than assuming all common causes are observed.
- Conditional-independence tests are underpowered in small samples (n < 100), producing spurious or missing edges.
- LiNGAM is undefined when its non-Gaussian assumption fails; PC must then be extended to FCI.
- With many variables the parent-set search becomes costly and unreliable unless its size is bounded.
- Recovered graphs require domain expertise to justify a genuinely causal reading.
Frequently asked
How is causal discovery different from regression?
Regression estimates the effect of predictors on an outcome whose causal role you have already assumed. Causal discovery instead tries to learn the direction of the arrows themselves — which variables cause which — directly from the data's (conditional) independence and distributional structure.
When should I use FCI instead of PC?
PC assumes there are no hidden confounders. FCI relaxes this and is appropriate when unmeasured common causes may exist; it returns a graph that explicitly marks where latent confounding could be present.
What does LiNGAM need that PC does not?
LiNGAM requires a linear model with additive, acyclic, non-Gaussian noise. In return, it can fully orient every edge of the DAG, whereas PC and FCI typically leave some edges undirected within an equivalence class.
Can I treat a discovered DAG as proven causation?
No. The algorithms recover structure consistent with the data under their assumptions, but the causal reading must be supported by domain expertise. A reasonably large sample (at least about 100 observations) is also needed for the independence tests to be trustworthy.
Sources
- Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation, Prediction, and Search (2nd ed.). MIT Press. ISBN: 978-0262194402
- Shimizu, S., Hoyer, P. O., Hyvärinen, A., & Kerminen, A. (2006). A Linear Non-Gaussian Acyclic Model for Causal Discovery. Journal of Machine Learning Research, 7, 2003-2030. link ↗
How to cite this page
ScholarGate. (2026, June 1). Causal Discovery Algorithms (PC, FCI, LiNGAM). ScholarGate. https://scholargate.app/en/causal-inference/causal-discovery
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- DAG Causal IdentificationCausal inference↔ compare
- Difference-in-DifferencesEconometrics↔ compare
- Instrumental Variables in Health ResearchHealth Economics↔ compare
- OLS RegressionEconometrics↔ compare
- Propensity Score MatchingResearch Statistics↔ compare