Published March 31, 2026 | Version v1
Journal article Open

CausalAudit: An Open-Source Framework for Partition Stability Analysis of Causal Graphs in Clinical and Pharmacological Research

Authors/Creators

Description

We present CausalAudit, an open-source Python framework that operationalizes the Partition Stability Index (PSI) methodology introduced by Nguyen, Trần, and Lê (2026) for auditing the robustness of causal graphs in clinical and pharmacological research.

The perspectival-relational account of causation (Kriger 2026) holds that causal structure is not observer-independent but jointly produced by the territory's undirected conditional independence skeleton and the bounded observer's variable partition, system boundaries, and temporal ordering. Nguyen et al. (2026) formalized this account by defining a causal compression cost functional and the PSI metric, proving that PSI-maximizing partitions align with the territory's conditional independence structure. However, no ready-to-use software tool implementing these constructs has been available.

CausalAudit closes this gap. The framework accepts a tabular clinical dataset and automatically: (1) generates families of perturbed variable partitions via four perturbation operators (coarsening, refinement, rotation, boundary shift); (2) learns the MDL-optimal directed acyclic graph (DAG) for each partition; (3) computes global and per-class PSI using the Structural Intervention Distance; and (4) produces a structured report classifying each edge in the reference causal graph as robust, moderately robust, fragile, or artifact based on its edge robustness score (ERS).

The paper describes the six-module architecture (DataIngestor, PartitionPerturber, DAGLearner, GraphDistanceCalculator, PSICalculator, ReportGenerator), introduces a projected SID extension for cross-partition comparison, defines the edge robustness classification schema, and specifies a rigorous evaluation protocol with four testable predictions — without fabricating experimental results. Three clinical use cases are developed: RCT generalizability assessment across populations, pharmacovigilance signal evaluation, and systematic review methodology.

This work directly addresses the problem identified by Murthy, Balakrishnan, and Subramaniam (2026) regarding the transferability of RCT-derived causal claims across populations with different coarse-graining structures.

Keywords: causal inference; partition stability; PSI; causal audit; MDL; clinical trials; observer-relative causation; Python framework; pharmacological research; compression; bounded observers; DAG; Bayesian networks

Related identifiers:

  • References: Kriger, B. (2026). Against Causation. https://doi.org/10.5281/zenodo.18851848
  • References: Nguyen, D.-M., Trần, T.-L., & Lê, H.-P. (2026). Compression Cost, Partition Stability, and the Engineering of Causal Graphs. Architectures of Inference, 3(2), 117–131.
  • References: Murthy, K., Balakrishnan, A., & Subramaniam, P. (2026). Compression, Causation, and the Sciences. Perspectives in Foundational Science, 4(2), 112–121.

License: CC BY 4.0

Resource type: Journal article

Publication date: 2026-03-28

Journal: Chulalongkorn Journal of Computational Science and Applied Epistemology, Vol. 12, No. 1, pp. 1–8

Files

causalaudit_paper.pdf

Files (82.9 kB)

Name Size Download all
md5:ff91dd452727e54def36a164c8e57862
82.9 kB Preview Download

Additional details

References

  • Ankan, A. & Panda, A. (2015). pgmpy: Probabilistic graphical models using Python. Proceedings of the 14th Python in Science Conference, 6–11.
  • Beckers, S. & Halpern, J.Y. (2019). Abstracting causal models. Proceedings of AAAI, 33, 2678–2685.
  • Chickering, D.M. (2002). Optimal structure identification with greedy search. Journal of Machine Learning Research, 3, 507–554.
  • Grünwald, P. (2007). The Minimum Description Length Principle. MIT Press.
  • Janzing, D. & Schölkopf, B. (2010). Causal inference using the algorithmic Markov condition. IEEE Transactions on Information Theory, 56(10), 5168–5194.
  • Kim, J. (1998). Mind in a Physical World. MIT Press.
  • Kolmogorov, A.N. (1965). Three approaches to the quantitative definition of information. Problems of Information Transmission, 1(1), 1–7.
  • Kriger, B. (2026). Against causation: A formal argument that causality is a compression artifact of bounded observers, not a feature of reality. Institute of Integrative and Interdisciplinary Research. https://doi.org/10.5281/zenodo.18851848
  • Murthy, K., Balakrishnan, A., & Subramaniam, P. (2026). Compression, causation, and the sciences: Practical implications of observer-relative causal structure. Perspectives in Foundational Science, 4(2), 112–121.
  • Nguyen, D.-M., Trần, T.-L., & Lê, H.-P. (2026). Compression cost, partition stability, and the engineering of causal graphs: Formalizing and empirically testing the perspectival account of causation. Architectures of Inference, 3(2), 117–131.
  • Pearl, J. (2009). Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge University Press.
  • Peters, J. & Bühlmann, P. (2015). Structural intervention distance for evaluating causal graphs. Neural Computation, 27(3), 771–799.
  • Price, H. (2007). Causal perspectivalism. In Price & Corry (eds.), Causation, Physics, and the Constitution of Reality. Oxford University Press.
  • Rissanen, J. (1978). Modeling by shortest data description. Automatica, 14(5), 465–471.
  • Russell, B. (1913). On the notion of cause. Proceedings of the Aristotelian Society, 13, 1–26.
  • Sharma, A. & Kiciman, E. (2020). DoWhy: An end-to-end library for causal inference. arXiv preprint arXiv:2011.04216.
  • Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation, Prediction, and Search. 2nd ed. MIT Press.
  • Woodward, J. (2003). Making Things Happen: A Theory of Causal Explanation. Oxford University Press.
  • Zheng, X., et al. (2024). causal-learn: Causal discovery in Python. Journal of Machine Learning Research, 25(60), 1–8.