acceptodds
Under review as a conference paper at ICLR 2027

CausaBridge: An Identification-Gated Causal Foundation Model for Transferable Effect Estimation

Abstract

Foundation models amortize inference across tasks, but causal effect estimation has a prerequisite absent from ordinary prediction: the requested effect must be identifiable under the declared causal model and available distributions before a numerical estimator is invoked. We introduce CausaBridge, an identification-gated causal foundation model. A typed task contract is compiled into a candidate identification certificate and independently verified; only an accepted point certificate can execute the contract-conditioned estimator. Verified bounds bypass the neural path, while non-identified, unsupported, or empirically unsupported queries return typed non-point results. We establish point-output soundness, noninterference of non-point branches, stability of registered functionals to nuisance error, conditional invariance under kernel-preserving adapters, and a task-distribution generalization bound. A prespecified common-target study evaluates 18 checkpoints on 256 unseen tasks (4,608 evaluations); diverse-mechanism pretraining reduces calibrated ATE error by 0.0180 relative to fixed-mechanism pretraining (95% crossed-bootstrap CI: 0.0117–0.0242; ) and satisfies a four-part decision rule specified before confirmatory evaluation. A 5,376-job comparison study and a matched 312-job selective evaluation on IHDP, ACIC, Twins, and Jobs produce 309 authorized point results and three typed refusals. On supported ACIC repetitions, CausaBridge reduces ATE error relative to CausalPFN by 0.0741 (95% repetition-bootstrap CI: 0.0419–0.1112). On nine authorized MIMIC-based semi-synthetic scenarios, CausaBridge obtains 0.1353 ATE MAE and 0.1756 root precision in estimation of heterogeneous effect (PEHE) across two independent seeds; adapter-only transfer updates 5.8% of parameters and reduces both errors relative to full fine-tuning under the prespecified protocol. In five-seed structural replay, the complete executor produces no false point outputs and makes no neural calls on non-point branches. A linked real-cohort audit separately converts inadequate treatment support into a typed non-point outcome. CausaBridge thereby unifies exact-query authorization, typed non-point execution, and reusable causal effect estimation in a single verified pipeline.

Primary area:
causal reasoning
Keywords:
foundation models, causal effect estimation, causal identifiability, transfer learning, structural causal models, selective prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Z8NGOpenReview ↗
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 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.

  1. 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.

  2. 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.

  1. 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].

  2. 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.

  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.

  1. Multimodality. MIMIC full/EHR-only/no-image ATE MAE is 0.1353/0.1302/0.1304, and scratch outperforms adapter transfer. We now separate support for different observation interfaces from an accuracy benefit due to images: this experiment demonstrates the former, while the ablation does not establish the latter.

This argument supports authorization for the stated rules under explicit assumptions. The numerical comparisons are evaluated separately.

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.

Reviewer_7P96OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present CausalBridge, a framework that separates causal identification from numerical estimation. A compiler generates a certificate based on the causal query, specified graph, assumptions, and available data. An independent verifier checks the certificate and determines whether point estimation is allowed. Otherwise, the framework returns bounds, refusals, or unsupported outcomes. The proposed framework is evaluated via benchmarks including IHDP, ACIC, Twins, and Jobs.

Critical strengths

I feel the strongest contribution of this work is turning causal identification and verification into a systematic framework that determines whether neural estimation can proceed, while allowing bounds or refusals as valid outputs.

Critical weaknesses
  1. My main concern is that the proposed framework does not demonstrate a sufficiently strong contribution to causal inference. The certificate verifies whether the analysis follows the stated causal assumptions, but it does not verify whether those assumptions are correct. An incorrect but internally consistent causal model could therefore pass verification and produce biased estimates. For identifiable queries, existing identification methods and estimators can already answer the same causal questions. For unidentified queries, the framework either refuses estimation or provides standard Manski bounds in limited settings. Neither introduces new causal capabilities. While a useful contribution does not necessarily require new identification results, the paper should demonstrate clear improvements in estimation, efficiency, or handling uncertainty about causal assumptions. The current results do not convincingly show these benefits. In particular, the ATE improvements on public benchmarks mainly come from the DR anchor, while the pretrained residual does not improve PEHE in Table 11. The framework may be useful as an auditable system for causal estimation. However, it remains unclear whether integrating a causal foundation model provides meaningful benefits over a simpler pipeline that combines the same verification rules with existing estimators. A direct comparison with such a pipeline is needed to better establish the contribution.

  2. Additionally, the experiment do not show a clear benefit from the pretrained foundation model. Equation 8 preserves the DR anchor’s ATE by construction, so the reported ATE improvements on public benchmarks cannot be attributed to the foundation model. Table 11 also shows no improvement in heterogeneous treatment effect estimation. On ACIC, CausaBridge + SEDRA achieves a root PEHE of 1.1284, compared with 1.1267 for the DR anchor and 1.1271 for SEDRA trained from scratch. On IHDP, the corresponding values are 2.4783, 2.4742, and 2.4622. Therefore, the pretrained model performs worse than both alternatives on these datasets. Similarly, on the MIMIC-based semi-synthetic tasks, training from scratch achieves lower ATE MAE (0.1214 vs. 0.1353) and root PEHE (0.1691 vs. 0.1756) than adapter-only transfer. I found those results show that the framework provides a conservative fallback mechanism and allows parameter-efficient reuse, but they do not demonstrate consistent accuracy gains from causal pretraining. Stronger comparisons are needed to show what the pretrained foundation model adds beyond the existing DR estimator or a model trained from scratch.

Questions for authors
  1. Where does the pretrained residual consistently improve over both the DR anchor and a matched from-scratch residual? Please provide paired estimates and intervals.

  2. Can you provide the empirical evidence of the benefit integrating causal foundation model over a simpler pipeline that combines the same verification rules with existing estimators?

3 comments

Thank you for asking what pretraining adds beyond DR and matched target training. In an additional study of a separately pretrained tabular branch, mean root PEHE is 1.8270 versus 1.9487 for trained random initialization and 2.1115 for same-gate DR. Both paired intervals are below zero. We report this revised procedure separately from the original experiments and distinguish estimation evidence from the verification layer's compatibility with existing estimators.

  1. Contribution beyond the DR anchor. Submitted Equation 8 (revised Equation 5) preserves the DR mean, so the relevant contribution of the learned component is heterogeneous-effect estimation at that fixed mean. For the additional IHDP study, we revised the adaptation/readout procedure to use full target adaptation followed by a ridge fit to treatment and outcome residuals on learned features (Sections 4.6 and 5.1). The original experiments are retained and reported separately.

On this panel, mean root PEHE is 1.8270 for the pretrained model, 1.9487 for trained random initialization, and 2.1115 for DR under the same verification rules. The observed reductions are 6.25% and 13.47%. Paired differences are -0.1217 [-0.2754,-0.0039] and -0.2844 [-0.4365,-0.1113], approximate nominal individual 97.5% intervals for two fixed contrasts. Both mean differences favor pretraining at each of the three fixed source-training seeds. The random-initialization comparison tests the contribution of source pretraining within the same architecture; the DR comparison tests the additional heterogeneity contribution at the preserved mean.

Both neural arms train all 48.7M parameters for 128 target updates, with identical target data, losses, optimizer and selection opportunities. Pretraining adds 1,024 synthetic source updates; the control starts from the saved pre-source initialization. Both use the same ridge fitting procedure, selecting among six penalties by support-data leave-one-out error with features fixed. This selects the readout penalty with features held fixed; feature learning itself uses all support labels. Query outcomes and potential outcomes enter scoring only.

The panel comprises 120 repetitions and three fixed source seeds: 337 of 360 runs passed the support checks, 23 returned refusals, and none failed numerically. The primary paired analysis uses the 101 repetitions supported at all three seeds, with seed differences averaged within repetition before 20,000 bootstrap resamples. This panel was fixed after the first 40-repetition study did not meet its joint criterion because the DR interval included zero. Source weights were reused, without pooling stages or stopping early. Earlier runs remain reported. The intervals concern fixed source models and reused IHDP covariates, rather than the full adaptive research process.

We agree that the original Table 11 did not establish a consistent benefit from the pretrained residual. We have revised its interpretation accordingly and clarified that the historical random-core control trained adaptation components with a frozen backbone. The added experiment uses a fully trained random-initialization control and addresses the paired comparison requested in your review.

  1. Same verification rules with existing estimators. We also used CausalPFN through the verification layer. In 12 execution tests, the supported back-door ATE case preserved standalone predictions exactly; the other 11 cases did not invoke its predictor. On the same IHDP evaluation, CausalPFN obtains root PEHE 0.5274, or 0.5884 after matching its mean to DR; source and target budgets differ. The T-learner obtains 1.8497 (ours minus T: -0.0226, paired interval [-0.1943,+0.1270]). These references motivate keeping the verification layer independent of the numerical estimator. Our controlled pretraining evidence comes from the matched random/DR comparisons above.

The methodological contribution is to make the conditions for the requested effect explicit and enforce them before point estimation. As with identification itself, this is conditional on the declared graph and assumptions; it does not establish their truth. DR and CausalPFN can therefore share the verification layer.

  1. MIMIC. At full target data, scratch training obtains ATE MAE/root PEHE 0.1214/0.1691, compared with 0.1353/0.1756 for adapter transfer; scratch is also better at 5%. We now describe this experiment as demonstrating parameter reuse with 5.8% trainable parameters. The parameter fraction is reported separately from accuracy and total computation. The additional IHDP evidence supports a pretraining benefit under that protocol, while the clinical results define an important boundary to its generalization.

We hope these paired comparisons answer your request for evidence beyond DR and a matched trained-random control. Their conclusion is specific to the revised IHDP procedure.

Thank you again for your review. Our response addresses your request for paired comparisons against DR and a matched trained-random control, distinguishing the additional tabular procedure from the original experiments. We would appreciate your assessment of this evidence and any remaining concern about the contribution of pretraining.

Thank you for addressing my concern. I raised my score accordingly.

Reviewer_8QqFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes CausaBridge which couples causal identification with amortized causal-effect estimation using a shared pretrained model. It accepts a query qq defining the target causal function, a contract κ\kappa that includes graph, variable domains, available data etc. Depending on identification and empirical-support checks, the framework returns a neural point estimate, a partial-identification bound, or a refused/unsupported result. The experiments evaluate transfer across tasks and downstream estimation, including the combination of the pretrained model with doubly robust estimation through SEDRA.

Critical strengths
  • I find CausaBridge an interesting and potentially practical framework. It combines a shared pretrained causal model with explicit identification checks based on the query and causal contract. It seems very useful to check whether the requested causal quantity is identified before returning a point estimate, and otherwise returning a bound or an explicit refusal/unsupported result.
  • The proposed CausaBridge+SEDRA model achieves improved causal estimation performance on ACIC, IHDP, Twins, and Jobs benchmarks.
Critical weaknesses

The paper would benefit from clearer explanations of several concepts. As currently written, I find the paper difficult to follow, partly because I am less familiar with the software terminology. My primary concerns are as follows:

  • The proposed method is restricted to a limited set of identification strategies. It does not support other important settings such as instrumental variable regression, proxy causal learning, or general longitudinal treatment regimes.
  • The registered partial identification seems restricted to binary outcome case (Lemma 3). Why is it restricted to binary case while the point estimates support continuous outcomes? Is it not possible to extend this to more general outcome spaces?
  • For Equation 9, the paper writes that Appendix C shows this inequality and Theorem 6 proves it. I think this is the result from Mohri et. al. Foundations of Machine Learning; see Theorem 3.3. The paper should cite this result explicitly and clarify the role of its application here.

Small things

  • Define σ^\hat{\sigma} in Equation 6 in the text.
  • Define EE and VV in line 327 for the O-notation.
  • Define the R\mathfrak{R} as Rademacher complexity for Equation 9.
Questions for authors
  • Why is the registered partial identification restricted to binary outcomes while the point estimates support continuous outcomes? Could the registered bounds be extended to other outcome spaces, such as continuous outcomes?
  • How easily could a practitioner extend the proposed framework to additional point-identification strategies, such as instrumental-variable methods or proximal causal learning, under their corresponding identifying assumptions? What changes would be required in the provided code?
5 comments

Thank you for recognizing the practical value of checking identification before estimation. We now verify continuous-outcome ATE bounds under a declared finite range for both potential outcomes, state the current identification scope and extension requirements explicitly, and explain the framework through an adjustment example. We also give the requested notation and attribution of the standard generalization bound.

  1. Bounds for continuous outcomes. Lemma 3 in the submitted paper describes a binary-outcome bound family. Continuous point estimation and partial identification use different mathematical objects, so support for the former did not automatically provide a verified bound for the latter. We have extended the registered bounds to outcomes with a declared finite range: for binary treatment, if Y(a) lies in [L,U] almost surely under both interventions, then ATE lies in [L-U,U-L]. The verifier independently checks this interval and returns it without calling the neural estimator (Appendix C.2.2). This is a conservative population bound based on the assumed outcome support; the observed sample minimum and maximum do not establish that assumption. Consistency and the validity of the declared range remain prerequisites.

We have also clarified the binary statement. The verified interval is [-1,1]. The narrower formula displayed in the submission additionally requires observed treatment-outcome probabilities, which the original checker did not verify. The revised statement distinguishes those probability-based bounds from the support-based intervals verified here, and the endpoint checks have been corrected accordingly.

  1. Extending identification. The implemented point rules cover marginal back-door, pure-DAG conditional back-door, restricted singleton front-door, and restricted source-to-target adjustment with explicit cross-domain invariance and support assumptions. We now state this scope directly; the implemented search is not a complete test of non-identifiability.

An extension to IV or proximal identification would add the appropriate identifying formula, checks of its assumptions and data requirements, and a compatible numerical solver. The same verification-before-estimation design would apply. The mathematical requirements depend on the estimand: IV assumptions need not identify the ATE considered here, while proximal identification may require bridge existence/uniqueness assumptions and a bridge-equation solver. Thus the reusable part is the workflow; each additional identification strategy needs its own mathematical and numerical components. This also explains why the present implementation focuses on the stated rule families.

  1. Clearer presentation. We now introduce the framework with an adjustment example. Given baseline X, treatment A and outcome Y, the compiler proposes an adjustment formula. The verifier checks whether X is admissible in the supplied graph, the required distributions are available, and the formula answers the requested effect. An accepted point formula permits estimation; bounds, refusals and unsupported results are explicit alternatives. The checks are conditional on the supplied causal graph and assumptions, whose substantive validity remains the analyst's responsibility.

  2. Generalization bound. Thank you for identifying the attribution issue. We now explicitly cite Mohri, Rostamizadeh and Talwalkar, Foundations of Machine Learning, Theorem 3.3. Submitted Equation 9 (revised Equation 6) applies that standard result to bounded losses and independent tasks from the declared distribution. Its role is to characterize task-distribution generalization under those assumptions. R_N denotes expected Rademacher complexity; the empirical counterpart conditions on the sampled tasks.

  3. Notation. V and E denote graph vertices and edges. In submitted Equation 6, s_i0 and s_i1 are predicted outcome log scales, sigma_i=sqrt(exp(2s_i0)+exp(2s_i1)), and sigma_ATE=sqrt(sum_i sigma_i^2)/n. We have added these definitions and distinguished these model-derived scales from calibrated causal confidence intervals.

We hope the worked example, explicit scope and bounded-continuous extension resolve the clarity and outcome-space concerns raised in your review.

Thank you again for your review. We would appreciate your assessment of our response on continuous-outcome bounds, the supported identification rules and the clarified definitions. Please let us know if any of these remains a concern for your recommendation as the discussion period comes to a close.

Dear authors, thank you for your responses and for improving the clarity. The mentioned edits address my comments on presentation, and I will increase my presentation score. However, the revised manuscript is now 40 pages (18 more pages than the original submission), and contains extensive edits, including corrections to the experimental results. I appreciate the explicit acknowledgement of these corrections in the manuscript(e.g. Appendix M). However, these edits are not clearly highlighted in the rebuttal, and it makes it difficult to reassess the paper as a whole. My overall rating therefore remains unchanged.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of responses and revisions

Dear reviewers and Area Chair,

Thank you for taking the time to revisit our work. The discussion helped us clarify what CausaBridge verifies before estimation and which conclusions the experiments support.

Reviewer 8QqF. The reviewer welcomed the clearer presentation and noted that the benchmark corrections were difficult to trace in our first response. We corrected the benchmark entries using the original predictions. On IHDP, the corrected ATE MAE is 0.2866 for CausaBridge+SEDRA and 0.1139 for CausalPFN. On Twins, the corrected CausaBridge+SEDRA results are 0.0210 ATE MAE and 0.3297 root PEHE. We withdrew the corresponding lowest-error claims. Revised Table 8 gives the corrected benchmark comparison, revised Table 9 gives the corrected component results, and Appendix M documents the changes. One finding that still holds in the original synthetic experiments is that diverse pretraining reduced calibrated ATE error by 0.0180 relative to fixed-mechanism pretraining. We also clarified the supported identification rules and the ATE bound for outcomes with a known finite range (Table 1 and Appendix C.2.2).

Reviewer 7P96. The reviewer asked what pretraining adds beyond the doubly robust (DR) estimator, since the original learned residual preserves the DR mean. We addressed this in a separate IHDP study (Section 5.1, Table 2). Across 101 matched repetitions, the separately pretrained tabular model achieved a mean root PEHE of 1.8270, compared with 1.9487 for matched training from random initialization and 2.1115 for DR under the same verification rules. CausalPFN achieved 0.5274 on this panel. Reviewer 7P96 confirmed that our response addressed their concern and raised their overall rating.

Reviewer Z8NG. The reviewer acknowledged the added argument for the four implemented identification rules (Appendix C.2.1) and asked what CausaBridge adds to ID followed by a compatible estimator. We agree that this combination suffices when the estimator evaluates the identified formula. CausaBridge makes that compatibility check explicit for each request. In Table 18, the supported back-door query reproduced CausalPFN’s standalone predictions exactly. The front-door effect was also identifiable, but the selected estimator could not evaluate its formula. CausaBridge reported that limitation before calling the predictor. These cases illustrate how a researcher can tell whether the obstacle lies in identifying the effect or in estimating it with the chosen model.

We appreciate your consideration of the corrected results, the additional matched study, and the discussion of the framework’s contribution.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.