1. BFCI — Best-Order Score Search + FCI Refinement
Type: Hybrid (score-based + CI tests)
Output: PAG
Knowledge: Fully supported (all Tetrad knowledge constraints)
Command: bfci
This algorithm was called BOSS-FCI through Tetrad 7.6.10 and is the same algorithm under a new name. The old name survives in the program only so that sessions saved by those versions still open; new work should use BFCI.
BFCI is marked experimental in the current release, so it is hidden from the algorithm chooser until you turn experimental methods on, and the description below may lag the code. See Experimental Methods.
BFCI is one of the mixed-strategy latent-variable algorithms introduced in
Ramsey, Andrews & Spirtes (2025). It combines:
BOSS — a score-based ordering search that produces a high-quality CPDAG, and
An FCI-style latent-variable correction stage, which transforms the CPDAG into a PAG.
The resulting algorithm is faster than GFCI and usually more accurate, particularly in medium- and high-dimensional settings.
1.1. Key Idea
BFCI follows the general X-FCI template:
Score phase
Run BOSS to produce a CPDAG (causal sufficiency assumed for this step).
BOSS evaluates local SEM-BIC–type scores under different orderings, using memoization and efficient updates.
Latent-variable correction
Treat the BOSS CPDAG as an “oracle skeleton” and apply:
Orientation from separating sets
Possible-D-SEP pruning
Zhang-style orientation propagation (Zhang, 2008)
PAG legality checks (maximality, acyclicity, almost-cycles, endpoint corrections)
This yields a PAG that accounts for latent confounders and (optionally) selection bias.
BFCI is therefore a hybrid algorithm: it uses score-based reasoning for adjacency decisions and Zhang-style constraint-based logic for completing a PAG.
1.2. When to Use
You want a PAG but pure FCI is too noisy or too slow.
BOSS performs well on your data (medium-to-large
p, moderate sparsity).You want a drop-in alternative to GFCI with better scalability.
You have mixed continuous/discrete data (BOSS supports mixed BIC scoring).
You want a method that scales to dozens–hundreds of variables with latent confounding.
1.3. Strengths
More accurate than GFCI in most regimes (Ramsey et al. 2025).
Fewer false positives and cleaner PAGs than standard FCI.
Leverages BOSS’s robust adjacency selection.
Parallelizable and efficient for larger variable counts.
Fully knowledge-aware (forbidden edges, required edges, tiers, temporal constraints).
1.4. Limitations
BOSS phase assumes causal sufficiency, so the initial CPDAG can still suffer mis-orientations in heavily confounded models.
Still requires CI tests for the correction phase (although fewer than GFCI).
For extremely dense graphs, score-based phases can slow down.
1.6. Prior Knowledge Support
BFCI fully supports knowledge.
You may connect a Knowledge box in the Tetrad GUI or provide a
Knowledge object programmatically.
Enforced constraints:
Required edges
Forbidden edges
Tier / temporal ordering
Forbidden ancestor/descendant relations
Any additional Tetrad structural constraints
Knowledge is enforced during both:
The BOSS adjacency/orientation decisions, and
The FCI-style PAG refinement.
1.7. Key Parameters in Tetrad
BFCI inherits parameters from BOSS and from the FCI refinement step.
CamelCase names (GUI / script API) shown.
1.7.1. BOSS-stage parameters
Parameter |
Meaning |
|---|---|
|
BIC penalty multiplier (higher → sparser). |
|
Maximum degree allowed during scoring. |
|
Parallelism level for scoring. |
|
Print decision logs. |
1.7.2. FCI-refinement parameters
Parameter |
Meaning |
|---|---|
|
Max conditioning set size for CI tests. |
|
Order-independent adjacency removal. |
|
Whether to disallow interpreting circles as selection bias. |
|
Print CI tests / orientations. |
(Exact list depends on the current Tetrad release.)
1.8. Reference
Ramsey, J., Andrews, B., & Spirtes, P. (2025).
Efficient Latent Variable Causal Discovery: Combining Score Search and Targeted Testing.
arXiv:2510.04263.
1.9. Summary
BFCI = BOSS CPDAG + FCI latent-variable correction.
A fast, accurate hybrid PAG learner that outperforms GFCI in most settings while remaining fully knowledge-aware and scalable.