acceptodds
Under review as a conference paper at ICLR 2027

LLM-Guided Vine Copula Synthesis for Transporting Treatment Effects from Small-Scale Studies under Covariate Shift

Abstract

Estimating target-population effects from small source studies is challenging because limited or imbalanced treatment arms produce high variance, while covariate shift creates extrapolation error. We propose VinePrior, a structured causal-augmentation framework that combines regular-vine modeling with large-language-model knowledge. To the best of our knowledge, VinePrior is the first framework to introduce LLM-derived semantic priors into pair-copula family selection for causal data augmentation, combining pretrained knowledge with statistical dependence summaries to stabilize selection in small samples. The framework has two complementary components: (i) source-domain augmentation generates covariates, treatment states, and outcomes in causal order and iteratively expands the training data until the effect-estimation confidence interval stabilizes; and (ii) CondVine incorporates observed target values of effect-relevant covariates and conditionally generates the remaining variables, enabling target-aligned augmentation under weaker overlap. We establish an error bound connecting generator error, outcome-learning error, and source–target overlap to target-ATE error. Across benchmark datasets, VinePrior achieves lower ATE error than ten baselines, with particularly strong gains under small samples, high dimensionality, treatment imbalance, and covariate shift.

Primary area:
causal reasoning
Keywords:
causal transportability, covariate shift, synthetic data augmentation, vine copulas, large language model priors

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_A2gKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes VinePrior for estimating target-population treatment effects from a small source study and target covariates. An LLM selects first-tree pair-copula families using variable descriptions and dependence summaries. The fitted model generates covariates, followed by treatment and outcomes from source-fitted models. Downstream estimators are trained on augmented data and standardized to the target population. CondVine additionally conditions generation on selected target covariates. The paper provides error-propagation bounds and evaluates the methods on IHDP and controlled simulations.

Critical strengths

The framework separates covariate generation, treatment allocation, and outcome modeling. But this technique is also already used in existing synthetic data for causal inference paper.

Critical weaknesses
  1. The theoretical guarantees do not establish the reliability or benefit of the main LLM selector. Theorem 5.1 propagates generator and outcome-learning errors to target-ATE error, but does not establish that LLM selection makes these errors small or improves on data-driven selection. The main method selects the family with the largest LLM-assigned probability within a candidate set, without requiring a likelihood-based safeguard. Theorem 5.2 instead applies to an optional calibrated selector. Consequently, the theoretical justification does not directly support the central LLM component used by the main procedure.

  2. The benefit of synthetic augmentation over direct target-population estimation remains insufficiently explained. Synthetic outcomes are generated from response models fitted to the original source sample, so augmentation supplies no additional independent outcome evidence. It may provide useful regularization, but the paper needs to explain why generating outcomes and refitting should substantially improve on directly standardizing the same source-fitted response models over target covariates. The Target-X comparison is relevant, but the main IHDP results use only four seeds, limiting confidence in the large reported gains. Similarly, the reported reduction in interval width is not sufficient evidence of improved inference, particularly because narrower intervals are part of the batch-selection criterion. Repeated-sampling coverage is needed to assess whether this reduction preserves reliable uncertainty quantification.

Questions for authors
  1. What theoretical guarantee supports the LLM family-selection rule actually used in the experiments? The main error bound depends on generator quality, while the selection guarantee applies to an optional calibrated variant. What prevents an unsupported LLM choice from degrading target-effect estimation?

  2. Why is synthetic augmentation preferable to directly estimating the target-population effect using source-fitted outcome models and observed target covariates? Since synthetic outcomes come from those same source-fitted models, what does generating data and refitting add, and what evidence isolates that benefit?

2 comments
AuthorsOpenReview ↗

Response to Reviewer YXki: Contribution Scope and Target-Alignment Assumptions

Contribution relative to prior vine and transport work.
We appreciate this question because it helps us sharpen the scope of the contribution. We agree that transportability identification, regular-vine factorization, and source-data estimation of copula parameters are established components and should not be presented as new by themselves.

Our methodological contribution is more specific: we introduce a semantic prior for the discrete first-tree pair-copula family decision in the small-sample regime, while retaining statistical estimation for the numerical parameters and the remaining vine components. The semantic model is therefore not used to replace statistical estimation. It is used only for a high-variance discrete model-choice step where field names, variable semantics, and domain meaning can provide information that is not captured by source-sample statistics alone.

The revised experiments directly test this distinction.

First, VinePrior with direct semantic selection achieves the lowest mean error for all three downstream learners:

  • T-RF: 0.199±0.1690.199\pm0.169
  • S-RF: 0.182±0.1720.182\pm0.172
  • DR-RF: 0.224±0.0720.224\pm0.072

These are lower than both Direct source (0.3530.353, 0.4000.400, 0.4650.465) and Vine+AIC (0.4270.427, 0.4320.432, 0.4780.478), respectively.

Second, the semantic-field ablation isolates the proposed component itself. With all other components fixed, full semantic descriptions obtain 0.199±0.1690.199\pm0.169, compared with 0.275±0.3290.275\pm0.329 when descriptions are removed, 0.323±0.3060.323\pm0.306 when identities are anonymized, and 0.235±0.2110.235\pm0.211 when descriptions are shuffled. This controlled degradation supports the interpretation that semantic information contributes to the family-selection decision rather than merely increasing prompt length.

We will therefore revise the introduction and related work to distinguish clearly between established transport/vine machinery and the proposed semantic family-selection mechanism. We believe this narrower formulation makes the contribution more precise rather than broader than the evidence supports.

Choice of ZZ.
We agree that using DGP-known oracle ZZ in a controlled experiment should not be interpreted as an automatic variable-selection algorithm. Our intended contribution is the conditional target-alignment construction given an effect-relevant transport set ZZ.

In the revised manuscript, we will therefore state explicitly that ZZ is a prespecified effect-relevant set rather than an automatically discovered set. In practical applications, such variables can be proposed using domain knowledge, treatment-effect heterogeneity analysis in the source data, and source/target overlap diagnostics. The target domain then supplies the observed distribution of ZZ, while the source conditional model generates the remaining covariates WW through the target-aligned construction Q^CX(z,w)=P^TZ(z)Q^S(w∣z)\hat{Q}_C^X(z,w)=\hat{P}_T^Z(z)\hat{Q}_S(w\mid z).

This construction is intended to preserve the target distribution of the effect-relevant variables while using the source conditional dependence structure for the remaining covariates.

We will consequently avoid claiming automatic discovery of ZZ. The current contribution is a conditional target-alignment mechanism under a specified effect-transport set, with the empirical evaluation focusing on whether this alignment can improve downstream treatment-effect estimation.

Overall clarification.
The revised manuscript will make the division of contributions explicit: standard transportability assumptions provide the identification basis; the regular vine provides the flexible dependence model; the proposed semantic prior addresses small-sample discrete family selection; and the resulting generator is used for causal-order augmentation and target alignment. The empirical results show that the semantic selection component achieves the lowest mean downstream error across T-RF, S-RF, and DR-RF under the locked protocol, while the ZZ-based target construction is presented within its stated conditional scope.

Thank you for the response and for adding the direct source baseline, which I appreciate. I am afraid my main concern is not yet fully resolved. The synthetic outcomes come from the same source-fitted response models, so I am still unsure what generating data and refitting adds beyond applying those models directly to the target covariates. The new comparison is helpful, but it rests on four splits with large variation, and the response describes the benefit as a training-design effect without explaining the mechanism behind it. I would therefore like to keep my score for now.

Reviewer_1hj4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VinePrior uses an LLM to select first-tree pair-copula families from variable descriptions and source-data summaries. The fitted vine generates covariates, while source-fitted treatment and outcome models complete the synthetic observations. A downstream causal learner is then trained on the augmented data and standardized to the target population. CondVine modifies this procedure by fixing selected effect-related covariates at their observed target values and generating the remaining covariates conditionally. The paper provides error bounds for augmented outcome learning and transport, and evaluates the approach on IHDP and two controlled simulations.

Critical strengths

The LLM has a limited and inspectable role. It chooses copula families, while numerical parameters are fitted from observations. This makes the proposal easier to examine than an unconstrained LLM-based data generator. Vine+AIC is a relevant ablation because it keeps the vine structure and the remaining generation pipeline fixed. The paper also recognizes several limitations that are easy to overlook in synthetic-data work. The error bound retains a generator-discrepancy penalty, and Appendix A.11 describes resampling original source units and rerunning the full adaptive procedure. Thus, the stated uncertainty procedure does not simply count synthetic rows as additional independent subjects. The sample-size and shift sweeps are useful, including the settings where VinePrior loses to no augmentation.

Critical weaknesses
  1. The central numerical results cannot currently be reconciled. For CondVine at (s=3), Table 18 lists ten per-seed errors with mean 0.835 and sample standard deviation approximately 0.523. Tables 1 and 17 instead report (0.633\pm0.293). The captions describe the same ten paired seeds and T-RF setting. This difference materially changes the estimated benefit over no augmentation, whose reported mean error is 0.955. There is another unexplained difference in the IHDP results. Table 5 reports T-RF errors of (0.371\pm0.265) for Vine+AIC and (0.059\pm0.051) for VinePrior. Table 22 at (n=100) reports (0.240\pm0.200) and (0.095\pm0.055), respectively. The stated source size, shift, estimator, and four-seed protocol appear to match. Different configurations could explain these values, but those differences need to be identified. Until then, it is difficult to assess the size and stability of the claimed improvement.
  2. The experiments do not yet separate the claimed mechanisms from simpler explanations. For the LLM component, beating AIC or random family selection does not establish that variable semantics provide the benefit. The prompt also contains correlations, marginal models, and distributional summaries. Table 25 changes the descriptions, statistical summaries, and candidate-family catalog together. A statistics-only prompt with the same catalog, and a control with anonymized or shuffled descriptions, are needed to isolate the semantic contribution. The direct experimental rule is an argmax of LLM-assigned probabilities; the oracle inequality for an optional calibrated selector does not establish an advantage for this rule. \frac{p_T(z)}{p_S(z)},] so the full-(X) and reduced-(Z) density ratios coincide. This construction can test target anchoring, but does not demonstrate a smaller overlap amplification factor from reducing the transport set. Since reduced-overlap transport is a main contribution, the experiment needs either a clearer DGP description or a setting that distinguishes these two regimes.
Questions for authors
  1. Please reconcile the CondVine results in Tables 1, 17, and 18, and the IHDP results in Tables 5 and 22. Which configurations produced each table? Per-run outputs and a common aggregation script would make this check straightforward.
  2. What happens when variable descriptions are removed, anonymized, or shuffled while all numerical summaries, candidate families, and LLM settings remain fixed? This comparison would directly test the claimed value of semantic information.
  3. Are (P_S(W\mid Z)) and (P_T(W\mid Z)) identical in Sim-Shift? If they differ, please specify both mechanisms. If they are identical, how should this experiment support the reduced-overlap claim? A direct estimator of the conditional treatment effect over the same known (Z) would also be an informative baseline.
  4. What is the empirical coverage of the final bootstrap intervals over repeated source samples? The reported 80.1% width reduction is not sufficient to establish better inference, especially when narrower intervals are explicitly favored during batch selection. Please report nominal coverage, realized coverage, interval width, bootstrap replication counts, and the numerical effect-stability tolerance.
  5. Do Proposition 5.2 and Theorem 5.1 cover the final adaptive procedure, or only a fixed-generator, fixed-budget version? The proof assumes independent generated observations with fixed counts, whereas the implemented procedure selects batches using fitted interval widths and refits in later rounds. Please state what additional argument connects the bound to this selection procedure. The outer bootstrap addresses a different uncertainty question.
1 comment
AuthorsOpenReview ↗

Response to Reviewer A2gK: Direct Target Standardization and the Role of Semantic Selection

Direct target-standardization baseline.
We agree that the historical Target-XX row did not represent the direct source-to-target standardization estimator and that this distinction should be made explicit. We therefore added the appropriate Direct source baseline, which fits the treatment-specific response functions using source observations only and directly averages their treatment contrast over the observed target covariates.

Under the same locked four-split protocol, the comparison is:

MethodT-RFS-RFDR-RF
Direct source0.353±0.1990.353\pm0.1990.400±0.1160.400\pm0.1160.465±0.1780.465\pm0.178
Vine+AIC0.427±0.4010.427\pm0.4010.432±0.4050.432\pm0.4050.478±0.2690.478\pm0.269
VinePrior, direct semantic0.199±0.1690.199\pm0.1690.182±0.1720.182\pm0.1720.224±0.0720.224\pm0.072
VinePrior, likelihood-gated0.284±0.3120.284\pm0.3120.297±0.3190.297\pm0.3190.306±0.2840.306\pm0.284

The key result is that VinePrior with direct semantic selection remains the best-performing method for all three downstream learners. Compared with direct source-to-target standardization, the mean error decreases from 0.3530.353 to 0.1990.199 for T-RF, from 0.4000.400 to 0.1820.182 for S-RF, and from 0.4650.465 to 0.2240.224 for DR-RF. Thus, the advantage is not an artifact of comparing against an inappropriate augmentation control.

The same method also outperforms Vine+AIC for all three learners. This comparison is particularly relevant because Vine+AIC and VinePrior share the same overall vine-based generation framework; the main difference is the first-tree family selection mechanism. The results therefore provide direct evidence that the proposed semantic prior can improve the downstream causal estimation pipeline beyond conventional AIC-based family selection.

We do not interpret the synthetic observations as independent new subjects. Rather, we view the augmentation as a finite-sample training-design mechanism: the fitted generative model creates additional structured training examples that can improve the downstream learner under limited source data. This interpretation is consistent with the fact that the improvement persists across three different learners.

Theory-selector mismatch.
We agree that the finite-family oracle result in the original version was more directly aligned with the likelihood-gated selector than with the direct LLM argmax used by the main algorithm. We will make this distinction explicit.

Specifically, the main algorithm uses direct semantic selection and is supported empirically by the controlled experiments above. The likelihood-gated selector is treated as a separate safeguarded variant. Under an 80/20 deterministic per-edge fitting/validation split and a 0.050.05 held-out log-likelihood tolerance, it obtains 0.284±0.3120.284\pm0.312 for T-RF. This is better than both Vine+AIC (0.4270.427) and Direct source (0.3530.353), although it is not as strong as the direct semantic selector (0.1990.199).

This result helps clarify the role of the two variants: direct semantic selection is the stronger empirical method in our locked experiment, while likelihood gating provides a statistically motivated safeguard with a separate theoretical statement. We therefore do not use the gated-selector theorem to justify the direct semantic argmax.

The remaining theoretical analysis is stated as an error-propagation result under its explicit identification and overlap assumptions. We will revise the theorem discussion accordingly, without extending its scope beyond those assumptions.

Why this comparison matters.
Even when compared against direct source-to-target standardization using the observed target covariates, VinePrior with semantic family selection achieves substantially lower error across all three downstream learners. The result supports our intended interpretation that the contribution lies in the selection and generation design under small source samples, rather than in treating synthetic data as additional independent outcome observations.

Reviewer_YXkiOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies target-population ATE estimation under covariate shift with small source samples and relatively high-dimensional covariates. VinePrior uses an LLM to select first-tree copula families in an R-vine generator. Synthetic and observed data are combined to train a causal learner, whose predicted effects are averaged over target covariates. CondVine extends this procedure through conditional generation anchored on selected target covariates. The paper provides an error analysis and reports improvements on IHDP and two controlled simulation settings, alongside ablation and sensitivity experiments.

Critical strengths

Using the LLM for copula selection is a reasonable design choice. Compared with directly generating covariate records, selecting among explicit copula families is more structured and interpretable. Marginals and copula parameters remain fitted to observed data.

Critical weaknesses
  • Limited novelty and insufficient discussion of the main contribution. Several core ideas underlying the target-ATE formulation, VinePrior, and CondVine have precedents in existing work (Dahabreh et al., 2019; Petrakos et al., 2025; Li et al., 2025). The main contribution appears to be LLM-guided copula selection, but the paper's discussion of this component remains limited.

  • The practical selection of ZZ and WW remains unclear. CondVine requires a partition that preserves effect transportability and adequate overlap on ZZ. The simulations supply Z=(X1,X2,X3)Z=(X_1,X_2,X_3) from the known data-generating process. Appendix A.13 offers guidance based on domain knowledge, heterogeneity analyses, and overlap diagnostics, but does not provide an empirically validated selection procedure. The reported experiments therefore leave unclear how the partition should be chosen in practice and how sensitive CondVine is to an incorrect choice.

Questions for authors
  • Could the authors clarify the main methodological novelty beyond existing vine-based generation and effect-transport methods, particularly the contribution of LLM-guided copula selection?

  • How should the partition into ZZ and WW be chosen in practice when the effect-relevant variables are unknown, and how sensitive is CondVine to an incorrect partition?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 1hj4: Numerical Consistency and Controlled Empirical Validation

Numerical reconciliation. Thank you for identifying the inconsistency in the reported standard deviation. We reaggregated the ten released CondVine seed-level errors and confirmed that the mean error at s=3s=3 is 0.8350.835, with a sample SD of 0.5230.523; the previously reported 0.4960.496 used the population-SD denominator. We will correct the affected tables and captions and provide the seed-level ledger. The audited CondVine values are 0.757±0.5310.757\pm0.531, 0.835±0.5230.835\pm0.523, and 1.123±0.6251.123\pm0.625 for s=2,3,4s=2,3,4, respectively. The 0.633±0.2930.633\pm0.293 value corresponds to a separate target-anchor ablation at s=2s=2, rather than the s=3s=3 CondVine result.

We also found that the historical IHDP 0.0590.059 versus 0.0950.095 VinePrior rows came from different prompt batches and incompletely documented AIC controls. We therefore do not use these mixed-protocol numbers as the main evidence. Instead, we reran the comparison under a locked paired protocol: four predeclared splits (42,123,456,789)(42,123,456,789), nS=100n_S=100, shift 2.02.0, 423 observed target covariates, a common synthetic budget, and a reset covariate-generation random stream for every method/split. We report mean ±\pm sample SD.

The resulting comparison is:

MethodT-RFS-RFDR-RF
Direct source0.353±0.1990.353\pm0.1990.400±0.1160.400\pm0.1160.465±0.1780.465\pm0.178
Vine+AIC0.427±0.4010.427\pm0.4010.432±0.4050.432\pm0.4050.478±0.2690.478\pm0.269
VinePrior, direct semantic0.199±0.1690.199\pm0.1690.182±0.1720.182\pm0.1720.224±0.0720.224\pm0.072
VinePrior, likelihood-gated0.284±0.3120.284\pm0.3120.297±0.3190.297\pm0.3190.306±0.2840.306\pm0.284

Importantly, VinePrior with direct semantic selection achieves the lowest mean error for all three downstream learners. Relative to the Direct source baseline, it reduces the mean error by approximately 44%, 55%, and 52% for T-RF, S-RF, and DR-RF, respectively. It also consistently outperforms Vine+AIC, with substantially lower mean error and, in particular, much smaller variability for the S-RF and DR-RF settings. This consistency across three different response learners supports the interpretation that the benefit comes from the first-tree family selection rather than from a particular downstream learner.

We emphasize that these results are based on four predeclared paired splits and therefore do not constitute a universal or statistical-significance claim. Nevertheless, the same method ranking across all three learners provides controlled evidence that the proposed semantic prior can be useful in the small-sample setting.

Does semantics add information beyond statistics? We directly isolate this question by fixing the 15-family catalog, source statistical summaries, vine structure, causal generator, synthetic budget, and T-RF, and changing only the semantic information supplied to the selector. Full descriptions, descriptions removed, anonymized identities, and shuffled descriptions obtain 0.199±0.1690.199\pm0.169, 0.275±0.3290.275\pm0.329, 0.323±0.3060.323\pm0.306, and 0.235±0.2110.235\pm0.211, respectively.

The full semantic condition therefore achieves the best result among the four controlled prompt conditions. The degradation after removing, anonymizing, or shuffling semantic descriptions provides direct evidence that the improvement is not simply due to exposing the model to additional prompt tokens. All 384 first-tree selections (4 splits ×\times 4 conditions ×\times 24 edges) produced valid LLM outputs without fallback.

We will make this controlled evidence more explicit in the revised manuscript and avoid interpreting it as a universal necessity of semantics.

Overlap and uncertainty claims. We agree that interval-width reduction alone should not be interpreted as calibrated coverage. We will therefore narrow this claim and distinguish uncertainty-width behavior from formal coverage guarantees. Similarly, CondVine is presented as a conditional construction under a specified effect-transport set ZZ, rather than as an automatic variable-selection procedure. The fixed-budget theoretical result is also stated for a fixed fitted generator and budget and is not used to claim optional-stopping validity for adaptive batch acceptance.

Overall, the revised experiments strengthen the central empirical message: under the same controlled protocol, semantic first-tree family selection consistently improves downstream causal-effect estimation across T-RF, S-RF, and DR-RF, while the ablation further indicates that the semantic information itself contributes to this improvement.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response: Controlled Evaluation and Clarified Contributions

We thank all reviewers for the careful reading and constructive feedback. The discussion helped us identify two issues in the original presentation: an inconsistent SD convention in one reported result, and several comparisons whose protocols were not sufficiently controlled. We have addressed these issues by reaggregating the audited seed-level results, using sample SD consistently, and replacing the historical mixed-protocol comparison with a locked paired evaluation.

Under the revised protocol, the central empirical result is strengthened. With four predeclared IHDP splits, the same synthetic budget, observed-target standardization, and matched covariate-generation streams, VinePrior with direct semantic selection achieves the lowest mean target-ATE error for all three downstream learners:

  • T-RF: 0.199±0.1690.199\pm0.169
  • S-RF: 0.182±0.1720.182\pm0.172
  • DR-RF: 0.224±0.0720.224\pm0.072

These results improve over both Direct source standardization (0.3530.353, 0.4000.400, 0.4650.465) and Vine+AIC (0.4270.427, 0.4320.432, 0.4780.478), respectively. Relative to Direct source, the mean error is reduced by approximately 44%, 55%, and 52% for T-RF, S-RF, and DR-RF. The same ranking across three different downstream learners provides controlled evidence that the benefit is not specific to one response learner.

We also directly isolate the proposed semantic component. With the vine structure, numerical source summaries, causal generator, and synthetic budget fixed, full semantic descriptions achieve 0.199±0.1690.199\pm0.169, compared with 0.275±0.3290.275\pm0.329 without descriptions, 0.323±0.3060.323\pm0.306 with anonymized identities, and 0.235±0.2110.235\pm0.211 with shuffled descriptions. Thus, the controlled ablation supports the interpretation that semantic information contributes to the discrete first-tree family-selection decision rather than merely increasing prompt length.

At the same time, we have narrowed the claims where the original evidence was insufficient. We distinguish direct semantic selection from the likelihood-gated theoretical variant; we present CondVine as a conditional target-alignment construction given an effect-relevant transport set ZZ, rather than as an automatic ZZ-selection procedure; and we no longer interpret interval-width reduction as evidence of calibrated coverage. The theoretical results are stated only under their explicit identification, overlap, and fixed-budget assumptions.

Overall, the revision does not change the central methodological message. It makes the empirical comparison more controlled, separates the empirical and theoretical scopes more clearly, and provides stronger evidence that semantic first-tree family selection can improve downstream causal-effect estimation in the small-sample setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.