Research map / Probabilistic and scientific modeling / Causal discovery
The third pillar of causal analysis? A measurement perspective on causal representations
Read the paperCheck your idea against 108,967 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.
Reinterprets causal representation learning as a measurement model and proposes T-MEX, a conditional-independence-test-based Hamming-distance score that evaluates whether learned representations exclusively measure the intended latent causal variables, validated on simulations and the ISTAnt ecological video benchmark.
Problem
Causal representation learning produces representations whose usefulness for downstream causal tasks (reasoning, discovery, ATE estimation) is unclear, and common CRL metrics (R^2, MCC, SHD) can give misleading assessments when latent variables are causally related.
Approach
Treat learned representations as proxy measurements of latent causal variables within a measurement model; check exclusivity of each measurement via conditional independence tests against the hypothesized measurement model, aggregating the test outcomes into the T-MEX score (Hamming distance between the adjacency matrix and test-rejection matrix).
Key results
- On simulations, Model A (correct measurement) gets T-MEX 0.12±0.33 versus 1.18±0.39 for Model B and 2.00 for corrupted Model C, while R^2 for Z1 is only 0.67 in B versus 0.998 in A.
- On ISTAnt across 2,400 classifier configurations, models with T-MEX=0 showed higher accuracy and lower ATE bias than T-MEX=1 models; Mann-Whitney U test p-value 0.0047.
- Under weak causal relations, T-MEX gives about 1.01 for an entangled representation while R^2 and MCC both give 1.0, misleadingly suggesting perfect identification.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the CRL model sufficiently rather than stopping at insufficient training (ablation) | CRL representation in the numerical simulation · Numerical simulation | T-MEX: $1.1800 \pm 0.3881$ -> $0.1200\pm 0.3283$ | Helped |
| Mix the identified representation with Z2 and Z3 (ablation) | Model A representation · Numerical simulation | T-MEX: $0.1200\pm 0.3283$ -> $2.0000 \pm 0.0000$ | Hurt |
| Mix Z1 and Z2 in the representation (ablation) | Perfect causal representation · Nonlinear SCM | abs(ATE bias): $0.1532 \pm 0.0283$ -> $0.5357 \pm 0.0540$ | Hurt |
| Use T-MEX to assess representation quality instead of relying on R² or MCC | Causal representation evaluation under weak causal relations · Weak causal relations | detection of entanglement: T-MEX: $1.0100 \pm 0.0995$; R² and MCC: $1.0000 \pm 0.0000$ | Helped |
| Add independent Gaussian noise to measurements (ablation) | T-MEX measurement evaluation · Noisy measurements | T-MEX: $0.0100 \pm 0.0100$ -> $0.3900 \pm 0.6497$ | Mixed |
| Use PCM with random forest instead of the original test choice | T-MEX evaluation · Numerical simulation | ranking of models by T-MEX: null | No clear effect |
| Use KCI instead of the original test choice | T-MEX evaluation · Numerical simulation | ranking of models by T-MEX: null | No clear effect |
| Use PCM (RF) instead of PCM for the model comparison | T-MEX evaluation of Model A versus Model B and Model C · Numerical simulation | T-MEX ranking: Model A $0.0000 \pm 0.0000$; Model B $0.8000 \pm 0.7559$; Model C $2.0000 \pm 0.0000$ | No clear effect |
| Use KCI instead of PCM for the model comparison | T-MEX evaluation of Model A versus Model B and Model C · Numerical simulation | T-MEX ranking: Model A $0.0400 \pm 0.1979$; Model B $0.1600 \pm 0.4218$; Model C $2.0000 \pm 0.0000$ | No clear effect |
| Compare representations with T-MEX = 1 versus T-MEX = 0 | ISTAnt representation models · ISTAnt | absolute ATE bias: p-value $0.0047$ | Hurt |
Surprising findings
- R^2 and MCC assign maximal or near-maximal scores to entangled representations under causal dependence, misleadingly suggesting perfect element-wise identification.
- In a simulation, SHD was zero in 77% of cases even though measurements were entangled, so causal discovery on learned representations can look correct despite poor identification.
Limitations
- T-MEX depends on the validity and power of the chosen conditional independence test, which is hard to guarantee for high-dimensional conditioning sets.
- Measurement model must be specified from identifiability theory, task assumptions, or ground truth knowledge; the framework does not handle misspecified latent causal models.
- Only exclusivity is tested nonparametrically; parametric constraints on measurement functions are not incorporated.
Experimental setup
Datasets: ISTAnt simulated linear and nonlinear SCM data
Benchmarks: ISTAnt
Metrics: T-MEX R^2 MCC Spearman correlation absolute ATE bias classification accuracy p-values SHD
Baselines: R^2 score mean correlation coefficient (MCC) structural Hamming distance (SHD) generalized covariance measure (GCM) kernel conditional independence test (KCI)
Builds on: causal representation learning identifiability (multiview block-identifiability) measurement models (Silva et al. style latent structure) projected covariance measure AIPW estimator DINOv2
Closest papers by meaning
- Unifying Causal Representation Learning with the Invariance Principle (ICLR 2025 · similarity 0.92)
- Identifiability Guarantees for Causal Disentanglement from Purely Observational Data (NeurIPS 2024 · similarity 0.89)
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.89)
- Causal Representation Learning via Counterfactual Intervention (AAAI 2024 · similarity 0.88)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.88)
- Local Identifying Causal Relations in the Presence of Latent Variables (ICML 2025 · similarity 0.88)
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.