38. PC-Max — PC with Maximum-p Collider Orientation
Type: Constraint-based
Output: CPDAG
Implements: Ramsey (2016)
Knowledge: Fully supported (same as PC)
PC-Max is an enhanced collider-orientation variant of the PC algorithm. It keeps the entire adjacency phase unchanged, but modifies the orientation of unshielded colliders by selecting the separating set with the maximum p-value among all valid separating sets. This greatly reduces false collider orientations and improves arrowhead precision—especially in high-dimensional or moderate-sample settings.
In Tetrad, PC-Max is the default configuration of PC.
38.1. Key Idea
PC-Max improves collider orientation while leaving all other PC operations unchanged.
Collect all separating sets S such that X ⟂ Z | S.
Compute p-values for each CI test.
Choose the separating set with the maximum p-value.
Orient X → Y ← Z iff Y ∉ S*.
All further orientations use standard Meek rules.
38.2. When to Use
You want higher orientation accuracy than standard PC.
Datasets are high-dimensional, noisy, or small-sample.
You want a drop-in replacement for PC with no extra settings.
38.3. Relation to Standard PC
Adjacency phase: identical
CI tests: identical
Orientation propagation: identical except collider selection
Knowledge support: identical
Output: CPDAG
See the PC page for details.
38.4. Prior Knowledge Support
PC-Max inherits full background knowledge support:
forbidden/required edges
temporal/tier constraints
partial orientations
All constraints are enforced consistently during adjacency removal and orientation.
38.5. Strengths
Markedly fewer false colliders
Higher arrowhead precision
No additional computation
No new parameters
Tetrad’s default and recommended method
38.6. Limitations
Same as PC: assumes DAGs without latent confounding
Faithfulness and suitable CI tests required
For latent confounders, use FCI/GFCI/BOSS-FCI/FCIT
38.7. Key Parameters in Tetrad
PC-Max adds no new parameters.
Enabled with:
colliderOrientationStyle = "maxP"
All other parameters match PC.
38.8. Reference
Ramsey, J. (2016).
Improving accuracy and scalability of the PC algorithm by maximizing p-value.
arXiv:1610.00378.
38.9. Summary
PC-Max = PC + maximum-p separating sets for collider identification.
A simple but powerful improvement that yields cleaner, more reliable CPDAG orientations.