Research map / Probabilistic and scientific modeling / Causal discovery
Causal discovery needs explicit epistemic standards
Harald Kugler, Leonard Henckel, Sebastian Weichwald
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that causal discovery lacks shared epistemic standards for comparing and discussing the untestable assumptions behind its methods.
Problem
Causal discovery has many algorithms but few principles for choosing between them; existing comparisons via identifiability, consistency, and benchmarks rest on often untestable assumptions about the data-generating process, with no shared standards for discussing them.
Approach
Argues that the field needs explicit epistemic standards for comparing, prioritizing, and discussing the assumptions underlying causal discovery methods, rather than relying on identifiability, consistency, and benchmark performance alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FLOP (MEC learner, random DAG draw) instead of DirectLiNGAM | FLOP vs DirectLiNGAM · ER graphs, p=50, n=1000, beta=1.4 | SHD | Improved |
| use FLOP instead of DirectLiNGAM on chains without colliders | FLOP vs DirectLiNGAM · chains, p=50, n=1000 | SHD | Hurt |
| increase average degree of ER graphs | FLOP · ER graphs p=50, beta=1.2 | SHD | Hurt |
| push noise distribution toward Gaussian (beta approaches 2) | DirectLiNGAM · ER graphs p=50, d=10 | SHD | Hurt |
| push noise distribution toward Gaussian | FLOP · ER graphs p=50, d=10 | SHD | No clear effect |
| increase lower bound of edge weights (more varsortable) | VarSortRegress · chains p=20, beta=1.2 | SHD | Improved |
| increase average degree | VarSortRegress / DirectLiNGAM · ER graphs p=50 | SHD | No clear effect |
| variance standardization (iSCMs) to remove varsortability | benchmark design · benchmark sampled models | faithfulness regime | Mixed |
Closest papers by meaning
- Challenges and Considerations in the Evaluation of Bayesian Causal Discovery (ICML 2024 · similarity 0.89)
- On Causal Discovery in the Presence of Deterministic Relations (NeurIPS 2024 · similarity 0.89)
- Strong and Weak Identifiability of Optimization-based Causal Discovery in Non-linear Additive Noise Models (ICML 2025 · similarity 0.88)
- Foundations of Testing for Finite-Sample Causal Discovery (ICML 2024 · similarity 0.88)
- On the Recoverability of Causal Relations from Temporally Aggregated I.I.D. Data (ICML 2024 · similarity 0.88)
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
