Research map / Probabilistic and scientific modeling / Treatment effects
Falsification of Unconfoundedness by Testing Independence of Causal Mechanisms
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.
Proposes MINT, a two-stage test that falsifies unmeasured confounding by testing whether treatment and outcome mechanism parameters are independent across multiple environments under the independent causal mechanisms principle.
Problem
Unconfoundedness, a key untestable assumption for treatment effect estimation from observational data, cannot be verified; the paper asks how to falsify it using multi-environment observational data.
Approach
Assume treatment and outcome mechanism parameters are sampled independently across environments (independent causal mechanisms). Unmeasured confounding induces dependence between the estimated mechanism parameters, which is tested with a Frobenius-norm covariance statistic and a bootstrap-permutation calibrated threshold, without conditional independence testing and without requiring transportabil
Key results
- MINT consistently outperformed HGIC in power when varying the number of environments K in the linear well-specified setting, and all methods controlled type I error without unmeasured confounding in that setting.
- MINT falsified unconfoundedness when treatment-mechanism parameters or the confounder distribution varied across environments, matching the theoretical non-degeneracy condition.
- On the Twins semi-synthetic data MINT achieved higher power than the transportability test and HGIC, and kept type I error below alpha=0.05 when all five confounders were observed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MINT instead of HGIC while varying the number of environments | MINT · Synthetic linear data with unmeasured confounding | falsification rate | Helped |
| Increase the number of environments | MINT and HGIC · Synthetic linear data with unmeasured confounding | falsification rate | Helped |
| Increase samples per environment | MINT and baseline methods except HGIC with KCIT · Synthetic linear data with unmeasured confounding | falsification rate | Helped |
| Increase the number of observed covariates | MINT and HGIC with KCIT · Synthetic linear data with unmeasured confounding | falsification rate | Hurt |
| Use Pearson partial correlation instead of KCIT for HGIC | HGIC · Synthetic linear data with unmeasured confounding | falsification rate | Helped |
| Use MINT instead of the transportability test or HGIC | MINT · Twins semi-synthetic dataset with unmeasured confounding | falsification rate | Helped |
| Use a misspecified degree-1 polynomial working model | MINT · Synthetic polynomial data without unmeasured confounding | false positive rate | Hurt |
| Increase polynomial model degree beyond the true degree (ablation) | MINT · Synthetic polynomial data with unmeasured confounding | power | Hurt |
| Increase samples per environment when using more flexible polynomial models | MINT · Synthetic polynomial data with unmeasured confounding | power | Helped |
| Bootstrap datasets during permutation-based threshold calibration (ablation) | MINT · Type I error control | Type I error rate | Helped |
| Allow mechanism parameters influencing treatment assignment or the unmeasured-confounder distribution to vary across environments | MINT · Synthetic data with unmeasured confounding | falsification rate | Helped |
| Use the transportability test when outcome mechanisms vary across environments without unmeasured confounding | Transportability test · Synthetic data without unmeasured confounding | false positive rate | Hurt |
Surprising findings
- The transportability test produced false positives without unmeasured confounding, plausibly because mechanism changes violated transportability.
- Misspecified working models inflated the false positive rate of MINT, and increasing polynomial flexibility reduced its power.
Limitations
- Requires known feature representations (up to permutation and scaling); misspecification inflates type I error.
- Power declines as model flexibility grows, and the approach depends on the plausibility of the independent causal mechanisms assumption.
- Theory is for linear/GLM-type models; the kernelized extension is only sketched and not implemented.
Experimental setup
Datasets: Twins birth data (USA 1989-1991)
Benchmarks: simulated multi-environment data semi-synthetic Twins data
Metrics: falsification rate type I error power
Baselines: transportability test HGIC with Pearson partial correlation HGIC with KCIT
Builds on: independent causal mechanisms transportability-based falsification hierarchical graph independence constraint (HGIC)
Closest papers by meaning
- Detecting hidden confounding in observational data using multiple environments (NeurIPS 2023 · similarity 0.89)
- Detecting and Measuring Confounding Using Causal Mechanism Shifts (NeurIPS 2024 · similarity 0.87)
- Automating the Selection of Proxy Variables of Unmeasured Confounders (ICML 2024 · similarity 0.86)
- A Non-parametric Direct Learning Approach to Heterogeneous Treatment Effect Estimation under Unmeasured Confounding (NeurIPS 2024 · similarity 0.85)
- Doubly robust identification of treatment effects from multiple environments (ICLR 2025 · similarity 0.85)
- Data Fusion for Partial Identification of Causal Effects (NeurIPS 2025 · similarity 0.85)
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.