ResearchPod Summary
Constraint-based causal discovery algorithms, such as PC and FCI, rely on conditional independence (CI) tests to prune edges from a fully connected graph and orient causal relationships. Because these algorithms do not use a global score to evaluate graph fit, the accuracy of the final causal structure is entirely dependent on the reliability of the underlying CI tests. This survey provides a comprehensive synthesis of CI testing, categorizing methods into six families: partial-correlation, contingency-table, regression, nearest-neighbor, kernel, and machine-learning-based approaches.
A critical contribution of this work is the mapping of test-level statistical properties to graph-level failures. The authors explain that Type I errors (false positives) lead to the retention of spurious edges, while Type II errors (false negatives) cause the premature removal of true edges. Furthermore, the loss of statistical power as the conditioning set size increases—a common issue in high-dimensional data—can prevent the algorithm from identifying the separating sets required for correct v-structure orientation. This cascading effect means that even small inaccuracies in CI testing can fundamentally alter the resulting causal graph.
The paper evaluates the availability and implementation of these CI tests across major R and Python libraries. By comparing features such as support for mixed-type data, parallelization, and handling of small-sample regimes, the authors provide a decision framework for practitioners. The survey emphasizes that selecting a CI test is not merely a technical choice but a design decision that dictates the transparency and reliability of the resulting causal model. It highlights that while many tests exist, few are robust to the mixed-type, high-dimensional, and small-sample settings frequently encountered in biomedical research.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.