GFCM: A Tail-Sensitive Mixed-Type Conditional Independence Test for Causal Discovery

Technical Summary: GFCM – A Tail-Sensitive Mixed-Type Conditional Independence Test for Causal Discovery

Problem Statement
Constraint-based causal discovery algorithms, such as PC and FCI, rely fundamentally on the accuracy of their conditional independence (CI) tests. Existing practical CI tests, including partial correlation and the Generalised Covariance Measure (GCM) and its variants (WGCM, PCM), primarily detect dependence through conditional covariance or the conditional mean. Consequently, they are structurally blind to dependence structures that exist purely in the variance (scale), non-monotone mean relationships, or the tails of distributions. This limitation is critical in domains like finance and climate science, where systemic risk and extreme events often manifest as joint tail dependence or scale shifts that covariance-based measures systematically underestimate.

While other tests exist that detect dependence beyond covariance (e.g., Kernel-based tests like KCI, nearest-neighbor mutual information, or spectral GCM), they face significant barriers to integration into causal discovery pipelines:

1. Scalability: Many are computationally prohibitive for the repeated querying required by the PC algorithm.
2. Mixed-Type Support: Few natively handle mixed continuous and categorical data without ad-hoc, non-invariant transformations.
3. Robustness in PC: Fast residual-based tests often fail to maintain calibration (Type I error control) when conditioning on heavy-tailed variables or as the conditioning set depth increases, leading to inflated false positive rates.
4. Directional Asymmetry: Some tests fail to detect edges when the causal direction interacts with the test's construction (e.g., mean-quantile cross-products).

Methodology: The Generalised Feature Covariance Measure (GFCM)
The authors propose GFCM, a regression-cost, mixed-type CI test designed to be robust within constraint-based discovery. GFCM extends the GCM template by replacing the single identity residual feature with a configurable set of residual features targeting the mean, scale, and tails.

• Feature Construction: For variables $X$ and $Y$ conditioned on $Z$, GFCM computes:
  • Mean Residuals: $e_A = A - \hat{E}[A|Z]$.
  • Centered Scale Residuals: $c_A = |e_A| - \hat{E}[|e_A| | Z]$. The paper proves that centering is necessary for Neyman orthogonality; the uncentered version is biased.
  • Quantile Indicator Residuals: $r_\tau(A; Z) = \tau - \mathbb{1}\{A \le \hat{Q}_\tau(A|Z)\}$, targeting specific levels of the conditional distribution (e.g., tails).
  • Mixed-Type Handling: For categorical variables, the set uses centered one-hot residuals.
• Block Aggregation: The test forms three blocks of products:
  1. Moment Block: $\{e_X e_Y, c_X e_Y, e_X c_Y\}$ (targets mean and scale).
  2. Quantile Orientations: $\{e_X r_\tau(Y)\}_\tau$ and $\{r_\tau(X) e_Y\}_\tau$ (targets tail dependence).
     Each block is reduced to a p-value via a multivariate GCM quadratic form (calibrated via asymptotic $\chi^2$ or sign-flip bootstrap). The final p-value is the Cauchy (ACAT) combination of the three block p-values.
• Nuisance Estimation: To ensure validity and calibration, GFCM employs a nonparametric nuisance estimation strategy. It uses a cubic spline basis with a number of knots growing with sample size ($K(n) \approx 10n^{1/5}$) on rank-transformed conditioning variables. This avoids the approximation bias of fixed parametric designs. The design also includes specific blocks (linear, interaction, and cross-fitted single-index) to handle heavy-tailed conditioners and non-additive means.
• PC Integration Fixes: The authors identify and repair two specific failure modes when running sensitive tests inside the PC algorithm:
  1. Directional Asymmetry (F1): A mean-quantile cross-product may yield zero signal for a specific edge orientation. GFCM resolves this by running both orientations ($X \to Y$ and $Y \to X$) and pooling evidence.
  2. Nuisance Fragility (F2): Fixed-knot splines fail on heavy-tailed conditioners. GFCM resolves this via the growing-knot spline on rank-transformed data.

Key Contributions

1. The Test Construction: A mixed-type, regression-cost CI test that targets location, scale, and tail dependence using a configurable feature set.
2. Theoretical Validity:
   • Proof that centering the scale residual is required for Neyman orthogonality.
   • Characterization of the detection class: The test is consistent against alternatives where the cross-covariance of the feature set is non-zero, and powerless otherwise.
   • Introduction of $\Phi$-faithfulness: A generalized faithfulness assumption linking the test's detection class to the graph structure. The authors prove that PC with GFCM consistently recovers the CPDAG of the graph defined by this detection class.
3. Algorithmic Fixes: Identification and correction of directional asymmetry and nuisance fragility specific to running tail-sensitive tests inside PC.
4. Empirical Benchmarking: A systematic evaluation on synthetic and semi-synthetic data demonstrating that GFCM is the only fast, mixed-type test that maintains calibration and power as sample size and conditioning depth grow.

Results

• Calibration: On heavy-tailed, heteroscedastic, and mixed-type nulls, GFCM maintains nominal size (approx. 0.05) across varying conditioning depths ($|S|$) and sample sizes. Competitors like FFCI, the partial copula test, and boosted GCM exhibit severe size inflation (reaching 0.8–1.0) under similar conditions.
• Power: GFCM detects scale and tail edges that covariance-based tests miss. In "skew edge" scenarios (where $X$ drives the skew of $Y$ while mean and variance are constant), GFCM retains high power as conditioning depth increases, whereas competitors (FFCI, BLITZ, RCoT) collapse toward chance levels.
• Causal Discovery: In PC discovery experiments on random DAGs:
  • At small sample sizes ($n=2000$), GFCM performs comparably to FFCI.
  • At large sample sizes ($n=10^5$), GFCM achieves the lowest Skeleton Structural Hamming Distance (SHD) among size-valid tests, recovering nearly all edges (recall 0.99). Competitors like FFCI lose precision due to false positives, while others fail to detect scale/tail edges.
• Semi-Synthetic Validation: On real financial data (EuStockMarkets innovations) and the Causal Chamber light tunnel, GFCM successfully recovers injected scale and skew edges that covariance-based methods fail to detect.

Significance and Claims
The paper claims that GFCM fills a critical gap by providing a fast, mixed-type, and well-calibrated CI test that is specifically hardened for constraint-based causal discovery. The authors emphasize that their contribution is not the novelty of detecting dependence beyond covariance (which is established in literature), but rather the construction that makes such detection usable within the PC algorithm.

Key claims include:

• Robustness: Unlike other fast tests, GFCM does not inflate false edges as sample size or conditioning depth increases.
• Targeted Sensitivity: It allows users to configure the test to target specific dependence structures (e.g., tails) without requiring expert knowledge of the underlying heteroscedasticity structure.
• Theoretical Grounding: The $\Phi$-faithfulness framework provides a rigorous identification guarantee for the graphs recovered by PC when using GFCM.
• Limitations: The authors modestly note that the tail advantage is demonstrated on simulated and semi-synthetic data, as no fully real benchmark exists with both known ground truth and heavy tails. They also acknowledge that the ACAT combination can be mildly anticonservative in extreme tails, suggesting future work on graph-level error control.

In summary, GFCM enables causal discovery algorithms to recover structural edges driven by scale and tail dependence, which are invisible to standard covariance-based methods, while maintaining the computational efficiency and statistical validity required for large-scale discovery.