FCIT: Targeted Causal Discovery

• FCIT is a causal discovery framework that uses targeted conditional independence testing and recursive blocking to effectively mitigate latent confounding.
• It leverages score-based initialization and selective CI testing to reduce false independence claims and improve computational efficiency.
• The method consistently produces well-formed Partial Ancestral Graphs (PAGs), outperforming classical FCI in high-dimensional and confounded settings.

FCI Targeted-Testing (FCIT) is an algorithmic framework for causal structure learning in observational data, designed specifically to address the deficiencies of traditional Fast Causal Inference (FCI) methods in the presence of latent variables and selection bias. Classical FCI performs exhaustive conditional independence testing on exponentially many subsets, which leads to problems such as spurious independence claims, unreliable edge orientations, and excessive computational cost. FCIT introduces targeted-testing—a path- and score-guided approach for conditional independence (CI) testing and edge orientation—that yields well-formed Partial Ancestral Graphs (PAGs) with improved precision, reliability, and scalability (Ramsey et al., 5 Oct 2025).

1. Motivation and Key Limitations of Classical FCI

The core challenge addressed by FCIT is the “repeated testing problem” of classical FCI algorithms. In classical FCI, CI tests are performed over all subsets of adjacents or the Possible-D-SEP set, resulting in numerous spurious independence claims (false positives) and, consequently, extra or missing edges in the estimated PAG. This undermines both the statistical and structural reliability of the inferred causal model, especially under latent confounding or finite-sample noise.

FCIT seeks to mitigate these limitations through two principal strategies:

• Target only the most informative CI tests by exploiting score-based initial graph search.
• Employ a recursive blocking mechanism, ensuring that CI testing is restricted to path-minimal separating sets.

This approach reduces unnecessary statistical tests and improves both the interpretability and statistical soundness of the learned structure.

2. Methodological Framework and Core Algorithm

FCIT operates in several stages:

a. Score-Based Initialization

A score-based search algorithm, typically Best Order Score Search (BOSS), is applied to obtain a high-scoring CPDAG. This CPDAG encodes skeleton and (some) collider structure.

b. Conversion and Preliminary Orientation

The CPDAG is transformed into an initial PAG by introducing circle endpoints where uncertainty exists and retaining reliable collateral orientations.

c. Recursive Targeted Testing via Path Blocking

The core innovation is the block_paths_recursively procedure. For any node pair $(x, y)$, the procedure:

• Iterates through all paths $p$ between $x$ and $y$ in the current graph.
• Constructs a candidate blocking set $B$ on the fly, adding nodes to $B$ as needed to block all paths recursively.
• Removes the edge $(x, y)$ when a minimal blocking set renders all paths m-separated.

Mathematical Guarantee:

Let $B$ be the returned blocking set, then the property

$$\mathrm{adj}(x, y) \quad \vee \quad \forall p: \mathrm{Path}(x, y, p) \Rightarrow \mathrm{Blocked}(p, B)$$

holds, where $\mathrm{Blocked}(p, B)$ means $p$ is (m-)blocked given $B$.

d. Discriminating Path Integration

The search for discriminating paths—essential for correct collider orientation—is incorporated into edge removal. Candidate colliders over these paths are hypothesized, and CI tests are run on associated sets; edge removal is followed by propagation of mandatory orienting rules.

e. Final Orientation with Zhang Rules

Efficient implementations of the Zhang orientation rules are applied repeatedly. This step involves shortest-path algorithms for all-circle paths and semidirected path searches, ensuring the PAG structural constraints (acyclicity, maximality) are always satisfied.

3. Comparison to Related Algorithmic Approaches

Method	Conditional Independence Testing Scheme	Structural Validity	Empirical Performance
FCI	Exhaustive all-subsets testing	May fail w/ cycles	Computationally expensive
BOSS-FCI	Score-based adjacency, exhaustive CI	Valid	Moderate precision, costly
LV-Dumb	No CI; DAG-to-PAG conversion	Not generally valid	Very fast, often accurate
FCIT	Targeted, recursive blocking, score-guided	Always valid PAG	High precision, efficient

FCIT outperforms exhaustive methods (FCI, BOSS-FCI) in adjacency and orientation metrics, and yields strictly better structural validity than LV-Dumb, which cannot properly orient bidirected edges and ignores latent confounders.

4. Empirical Evaluation: Simulations and Real Data

• Simulated Data:

FCIT was run on datasets with 20–200 nodes and latent confounding. It achieved higher adjacency and arrow path precision than BOSS-FCI and GRaSP-FCI and required fewer CI tests, demonstrating scalability.

• Structural Validity:

The PAGs produced by FCIT are always structurally well-formed (no cycles, no almost-cycles, no violations of maximality). Competing methods may return invalid structures under noise or high-dimensionality.

• Real Data Example:

Applied to the Algerian Forest Fire dataset (after exclusion of deterministic features), FCIT correctly inferred the directed relationships: Fire is directly caused by Temperature and Relative Humidity, with Rain as an additional antecedent, and exogenous factors Month and Region.

5. Guarantees, Limitations, and Parameter Considerations

FCIT guarantees proper edge-minimality and well-formedness of the output PAG due to its recursive blocking and separation strategy. The path-based targeted testing reduces over-conditioning and thereby false independence claims.

Notable limitations and open issues include:

• Selection of penalty discounts for score search and $\alpha$-levels for CI tests must be tuned appropriately.
• While structural validity and empirical precision are empirically robust, theoretical completeness with respect to all possible latent-variable scenarios is not formally established for highly non-faithful data.
• Extension to time-series or dynamic settings and further automation of parameter tuning are proposed future research directions.

6. Significance for Causal Discovery and Impact on Practice

FCIT provides an efficient and principled solution to latent-variable causal structure learning. By focusing on the most informative conditioning sets and integrating score guidance with recursive path blocking, it improves both the reliability (reduced false positives/negatives) and tractability of causal discovery in high-dimensional and confounded settings. The method is suitable for applications ranging from scientific inference to large-scale biomedical and social science data, where latent confounding and selection bias are present.

FCIT represents a methodological advance over both exhaustive CI-test-based FCI and heuristic score-based approaches, achieving a balance between formal correctness and scalability (Ramsey et al., 5 Oct 2025). Its use of targeted path-blocking offers a template for further innovations in structure learning algorithms for complex causal systems.