CausalGuard: Conformal Inference Under Graph Uncertainty
CausalGuard, a novel approach, tackles the issue of graph uncertainty in causal inference through a structure-weighted conformal framework. This method combines graph-conditional doubly robust pseudo-outcomes, utilizing candidate directed acyclic graphs (DAGs) suggested by an edge prior from a large language model (LLM), which are then refined using conditional-independence tests and reweighted according to the Bayesian Information Criterion. A composite nonconformity score is employed to calibrate the posterior-weighted pseudo-outcome, ensuring distribution-free finite-sample marginal coverage. Its conditional mean coverage is assured under conditions of causal identification, overlap, stability of conditional-mean nuisances, and focus on valid adjustment strategies aligned with targets. This technique is detailed in a paper available on arXiv (2605.21928).
Key facts
- CausalGuard is a structure-weighted conformal framework for causal inference.
- It calibrates after aggregating graph-conditional doubly robust pseudo-outcomes.
- Candidate DAGs are proposed from an LLM-derived edge prior.
- DAGs are pruned by conditional-independence tests and reweighted by BIC.
- A composite nonconformity score calibrates the posterior-weighted pseudo-outcome.
- It provides distribution-free finite-sample marginal coverage.
- Conditional mean coverage is guaranteed under specific assumptions.
- The paper is available on arXiv with ID 2605.21928.
Entities
Institutions
- arXiv