Choosing an Algorithm

A curated guide to the algorithms most Tetrad users should start with.

Tetrad provides many structure-learning algorithms. This short list highlights the most recommended ones, with crisp summaries and links to full per-algorithm pages.

If you’re new to Tetrad, start here.


πŸ” Choosing an Algorithm

Most users choose between:

  • DAG/CPDAG algorithms (assume no hidden confounders)

  • PAG algorithms (allow hidden confounders and selection bias)

A quick rule of thumb:

  • No hidden confounders β†’ Use DAG/CPDAG methods
    (FGES, BOSS, PC, PC-Max)

  • Hidden confounders possible β†’ Use PAG methods
    (FCI, GFCI, BFCI, FCIT)



πŸ” DAG / CPDAG Methods (No Latent Confounders)

PC β€” Peter–Clark Algorithm
πŸ” Constraint-based β€’ πŸŽ›οΈ Ξ±-controlled β€’ Output: CPDAG
Removes edges using CI tests and orients them using collider and propagation rules. A classic method offering explicit statistical control via Ξ±.
β†’ Full page: PC

PC-Max
πŸ” Constraint-based β€’ πŸŽ›οΈ Ξ±-controlled β€’ Output: CPDAG
Variant of PC that selects separating sets with maximum p-value, improving orientation accuracy.
β†’ Full page: PC-Max


FGES β€” Fast Greedy Equivalence Search
πŸ“ Score-based β€’ ⚑ Highly scalable β€’ Output: CPDAG
Greedy forward–backward search optimizing BIC. Very fast and scalable; a strong general-purpose default.
β†’ Full page: FGES

BOSS β€” Best Order Score Search
πŸ“ Score-based β€’ 🎯 Order-based β€’ Output: CPDAG
Searches over variable orders and uses Grow–Shrink Trees to score efficiently; often produces sharper orientations than FGES.
β†’ Full page: BOSS


πŸŒ€ PAG Methods (Hidden Confounders Allowed)

FCI β€” Fast Causal Inference
πŸ” Constraint-based β€’ 🧩 Latent-capable β€’ Output: PAG
Extends PC with additional pruning and orientation rules (R0–R10) to represent latent confounding and selection bias correctly. The canonical PAG discovery algorithm.
β†’ Full page: FCI

GFCI β€” Greedy Fast Causal Inference
πŸŒ€ Hybrid β€’ 🧩 Latent-capable β€’ Output: PAG
Uses FGES to obtain a CPDAG, then upgrades it to a PAG via FCI-style pruning/orientation. Parent design for newer hybrids.
β†’ Full page: GFCI

BFCI
πŸŒ€ Hybrid β€’ 🧩 Latent-capable β€’ Output: PAG
Replaces FGES with BOSS, producing a sharper starting point and improved pruning/orientation accuracy. Excellent general-purpose PAG learner.
β†’ Full page: BFCI

FCIT β€” FCI with Targeted Testing
πŸŒ€ Hybrid β€’ 🧩 Latent-capable β€’ 🎯 Targeted CI tests β€’ Output: PAG
Uses score-guided selective testing to avoid low-value CI tests, improving stability and accuracy while guaranteeing a legal PAG. Often outperforms both FCI and GFCI.
β†’ Full page: FCIT


πŸ”§ Other Useful Algorithm Classes

Several other methods specialize in:

  • Orientation only (FASK, LOFS, skew-based methods)

  • Non-Gaussian structure (LiNGAM, ICA-based models)

  • Time-series (PCMCI)

  • Markov blankets

  • Deterministic relations

These are powerful in the right contexts but are not typical starting points.

β†’ See full catalog: Full Algorithm List


πŸŽ› Choosing CI Tests & Scores (Quick Guide)

  • Continuous (Gaussian-ish): Fisher Z, SEM-BIC

  • Discrete: G-test or Chi-square; BDeu/BIC

  • Mixed or nonlinear: KCI / RCIT (slower), basis-function methods (scalable)

  • Covariance-only datasets: Use methods accepting covariance matrices (e.g., BOSS, FGES)


⚠️ Common Pitfalls and Fixes

  • Graph too dense: lower Ξ± (PC/FCI) or increase penalty (FGES/BOSS)

  • Graph too sparse: raise Ξ± or decrease penalty

  • Odd orientations: try PC-Max, BOSS, or minimal prior knowledge

  • Slow runtime: limit depth; use RFCI or FCIT; increase threads