9. Conservative PC (CPC) β Conservative Collider Orientationο
Type: Constraint-based (conservative variant)
Output: e-pattern (equivalence class of CPDAGs)
Conservative PC (CPC) is a modification of the PC algorithm designed to avoid false-positive collider orientations by requiring stronger evidence before declaring a triple a collider. Instead of orienting all unshielded colliders using a single separating set, CPC checks all relevant separating sets and only orients a collider when all such tests agree.
This prevents erroneous collider orientations when CI tests are noisy or unstable.
9.1. Key Ideaο
Given an unshielded triple:
X β Y β Z with X not adjacent to Z
PC orients this as a collider X β Y β Z if Y β sepset(X, Z).
CPC strengthens this:
9.1.1. Conservative Collider Ruleο
For every separating set S such that
X β Z | S and S is a subset of adj(X) βͺ adj(Z):
If Y β S for all such S, conclude:
X β Y β Z (collider)If Y β S for all such S, conclude:
X β Y β Z (noncollider)Otherwise:
Leave the triple unoriented (ambiguous)
The resulting graph is an e-pattern rather than a CPDAGβsome orientations remain intentionally unresolved.
9.2. When to Useο
Use CPC when:
You expect small sample sizes, noisy CI tests, or unstable separating sets
False-positive collider orientations would be costly (e.g., downstream adjustment sets)
You want a more conservative, robust version of PC
CPC is strictly more conservative than PC-Max and classic PC.
9.3. Prior Knowledge Supportο
CPC fully supports background knowledge, including:
Required edges
Forbidden edges
Tier/temporal constraints
Any other
Knowledgebox logic in Tetrad
All constraints are respected during adjacency and orientation phases.
9.4. Strengthsο
Greatly reduces false-positive collider orientations
Robust under sampling variability
Produces a safe e-pattern that captures uncertainty
Still quite fast in practice
9.5. Limitationsο
May leave many triples unoriented (intentionally)
Produces an e-pattern rather than a CPDAG, so fewer directions may be implied
Conservative nature may propagate to downstream inference
9.6. Key Parameters in Tetradο
CPC is implemented as a collider orientation style inside the PC search wrapper, so it shares PCβs parameters:
Parameter (camelCase) |
Description |
|---|---|
|
Use stable (order-independent) adjacency search. |
|
Set to βConservativeβ to use CPC rules. |
|
Allow temporary bidirected edges. |
|
Maximum conditioning set size. |
|
FDR-controlled CI testing option. |
|
Lag structure for time-series data. |
|
Replicate lag structure across slices. |
|
Log CI tests and orientation steps. |
See the main PC documentation for general parameter behavior.
9.7. Referenceο
The original peer-reviewed publication:
Ramsey, J., Zhang, J., & Spirtes, P. (2006).
Adjacency-faithfulness and conservative causal inference.
In Proceedings of the 22nd Conference on Uncertainty in Artificial Intelligence (UAI-06), pp. 401β408.
A later archival version:
Ramsey, J., Zhang, J., & Spirtes, P. L. (2012). Adjacency-faithfulness and conservative causal inference. arXiv:1206.6843.
9.8. Summaryο
Conservative PC provides a safe, cautious variant of PC, orienting only those colliders supported unanimously by all relevant separating sets. It is ideal when avoiding false orientations is more important than maximizing orientation coverage.