- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes a framework that combines causal identification with a pretrained foundation model for treatment effect estimation. Given a causal query, an assumed causal graph, and the available data, the proposed method first checks whether the requested effect is identifiable using existing criteria. A separate verifier checks the proposed identification certificate before applying the neural estimator. If the query is not supported, the system returns a bound or a refusal instead of a point estimate.
For effect estimation, the authors pretrain a foundation model on synthetic SCMs and propose a doubly robust adaptation method that combines a traditional doubly robust estimator with the foundation model. The paper also provides theoretical results on the identification verification procedure and estimation error.
The experiments are conducted on synthetic tasks and several causal inference benchmarks. The results show that diverse-mechanism pretraining improves performance on synthetic tasks and yields competitive estimation performance on several benchmarks.
Critical strengths
The paper aims to address an interesting and practical problem: some data modalities can improve outcome estimation but are not admissible for certain causal queries.
Critical weaknesses
-
The paper is relatively hard to follow because many terms are not well defined or well explained in the paper. Examples include: (l67) "identification control plane", (l77) fail-closed interface, (l199) compiler version. The paper's presentation would improve significantly if it explained the technical terms and concepts more clearly.
-
Many of the terms in the main paper are not defined. For example, the loss terms in Equation 5. Also, a mathematical definition of the compiler and verifier would help improve the paper's overall presentation.
-
The paper claims a broad capability for the proposed framework, but only demonstrates it in relatively limited cases. In particular, the identification component only supports a few standard identification criteria, and the experiments do not sufficiently demonstrate the framework's effectiveness in more complex causal settings.
-
Theorem 1 seems to follow directly from the assumed soundness of the verifier and the fail-closed execution design. However, it is unclear how the verifier's soundness can be guaranteed. The theorem appears largely true by construction rather than providing a theoretical guarantee. Similarly, the theoretical guarantees of Theorems 2, 3, and 4 seem reasonable under the proposed assumption, but it is unclear how they are useful in practice. I would be happy to discuss this with the author if I misinterpreted these.
-
It is not clear whether the proposed framework improves the estimation. For example, Table 11 shows that the proposed method's performance is close to the DR anchor. So the main improvement over the baseline comes from the estimator's DR property, which is well known in the literature.
-
One of the main motivations of the paper is the multimodal learning setting; however, the experimental results do not demonstrate the benefit of multimodal learning. In Table 2, the EHR-only and no-image versions perform similarly to the full model. It is unclear whether the proposed multimodal framework offers a meaningful advantage over simpler alternatives.
Questions for authors
-
Could the authors clarify the advantage of the proposed verification method? If the causal graph is given, the identifiability of a causal query can be determined with the ID algorithm, which is sound and complete. So it is unclear what the verification part of the proposed framework contributes.
-
Apart from the identifiability result, could the authors provide a clear comparison to show the advantage of the proposed method over existing amortized causal inference methods, as well as a comparison of the proposed work with existing meta-learners for causal inference?
5 comments
Thank you again for your review. Our response discusses the rule-level verification argument under explicit assumptions and the distinction between identification and checking the current request. We would appreciate your assessment of whether a specific soundness concern remains within the stated rule scope.
Thanks for the detailed response. My main concern is about the significance of the paper's contribution. I appreciate the clarification of the verifier's role and the additional soundness argument. However, it is still unclear what substantive advantage the proposed framework provides over the identification algorithm coupled with a compatible estimator. For example, one can use the ID algorithm to check the identifiability without estimating the distribution. They apply estimation methods such as CausalPFN.
Overall, the revisions clarify what the framework does and under what assumptions, but do not yet sufficiently address why it represents a significant advance over existing approaches.
Thank you for raising the verifier-soundness question. We now give a case-by-case argument for the four implemented point-rule families under explicit premises (Appendix C.2.1). The new estimation comparisons use a separately pretrained tabular branch.
Relation to ID. ID derives a formula under a given causal graph. Our verifier checks the candidate against the current query, variable roles and available distributions before estimation. For example, a marginal ATE certificate cannot approve CATE at C: the formula must retain C and average remaining adjustment variables conditional on C. Changing the label is insufficient. These checks can follow an ID implementation or precede a conventional estimator. The contribution is the checked connection between identification and execution, with point, bound, refusal and unsupported outputs.
Verifier soundness. The execution guard establishes when an estimator may be called; causal validity also needs a rule-level argument. We now complement Theorem 1, which assumed soundness as A1, with a case-by-case argument linking the implemented marginal/conditional back-door, restricted singleton front-door and restricted transport conditions to their standard identification formulas (Appendix C.2.1). It requires causal Markov and modular intervention assumptions, consistency, positivity, correct distribution interfaces and finite moments for means. Graph and serialization routines remain trusted; the argument concerns this finite rule family.
The proof also clarified the transport premises. Compiler and verifier now require conditional potential-outcome-law invariance given X,S and source coverage of target baseline support, alongside consistency/positivity. Previously, a case with source ATE 0.2 and target ATE 0.6 could be accepted. Missing declarations are now rejected; their truth remains assumed. These premises justify integrating source conditional outcomes against the target baseline distribution. Historical records are retained.
Practical meaning and scope. Theorems 2–4 describe how errors in the specified plug-in kernels affect estimation and what an adapter must preserve. Their conditions are bounded kernels for stability, bounded predictions and independent held-out data for concentration, and preserved kernels, measures and roles for invariance. These specify when the guarantees apply, rather than follow automatically from learning a representation. General ID, IV, proximal and longitudinal rules lie beyond this implementation; unsupported does not prove non-identifiable. For binary treatment, a declared finite potential-outcome range [L,U] also permits the continuous-outcome support bound [L-U,U-L].
Definitions. The compiler proposes a formula from a query and contract; the verifier checks it against the query, declared assumptions and available distributions. “Control plane” names these checks; “fail-closed” means failed checks prevent point estimation; “compiler version” identifies compatible outputs. Revised Equation 3 specifies losses, weights, masks and reductions, distinguishing auxiliary CATE targets from the requested task-level effect. Revised Equation 6 cites Mohri et al., Theorem 3.3.
Estimation comparisons. The additional IHDP study uses revised adaptation/readout, with original experiments retained separately. Root PEHE is 1.8270, versus trained random initialization at 1.9487 and same-gate DR at 2.1115. Paired differences are -0.1217 [-0.2754,-0.0039] and -0.2844 [-0.4365,-0.1113], approximate nominal individual 97.5% intervals. Both neural arms receive 128 target updates and the same readout; pretraining adds 1,024 source updates. The analysis uses 101 matched repetitions out of 120 across three fixed source seeds. The panel was fixed after an earlier joint criterion was not met, without pooling or early stopping; earlier runs remain reported. Intervals concern fixed source models and reused covariates, rather than the preceding adaptive search. The predicted query-cohort mean remains fixed to DR, isolating heterogeneity estimation.
CausalPFN and S-, T- and X-learners obtain root PEHE 0.5274, 2.7257, 1.8497 and 2.7619. For T, the paired difference is -0.0226 [-0.1943,+0.1270]. Source and target budgets differ across estimators. We also ran native CausalPFN through the verification layer, preserving standalone predictions in the tested back-door case. An existing predictor can thus retain its numerical output while using the same request checks.
This argument supports authorization for the stated rules under explicit assumptions. The numerical comparisons are evaluated separately.