acceptodds
Under review as a conference paper at ICLR 2027

Group-Adversarial Transfer for Causal Effect Estimation: Aligning Domains without Compromising Target-Domain Causal Information

Abstract

Estimating average treatment effects (ATEs) is challenging under target-domain data scarcity. Although related source domains can provide additional information, distributional shifts may cause negative transfer. Effective transfer therefore requires reducing cross-domain discrepancies while preserving treatment-effect-relevant information and the treatment–control structure. We propose GADANet, a transfer causal learning framework that jointly learns representations for causal effect estimation and domain alignment. GADANet decomposes the learned representation into domain-alignment and treatment-effect subspaces and uses group-wise adversarial training to align source and target representations independently within treated and control groups. This design reduces cross-domain discrepancies while preserving treatment-effect-relevant information and avoiding cross-treatment alignment. We evaluate GADANet across eight simulation settings with varying target-domain sample sizes and distribution shift magnitudes. GADANet consistently achieves lower ATE errors, reducing error by up to relative to Baseline, relative to Warmstart, and relative to Pooling. Ablation studies confirm the contributions of both proposed components.

Primary area:
causal reasoning
Keywords:
Average treatment effect estimation, Transfer causal learning, Domain adaptation

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_JPfDOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the challenge of limited target-domain samples and covariate distribution shifts between the source and target domains. It decomposes the learned representations into domain-invariant information for domain-alignment subspace and treatment-effect subspace, and performs domain-adversarial alignment separately within the treatment and control groups to mitigate negative transfer.

Critical strengths
  1. The paper addresses a practical problem in causal transfer learning, where target-domain samples are scarce while source-domain samples are abundant, and proposes a method to tackle this challenge.

  2. The method is model-agnostic and can be conveniently integrated into neural causal models with shared representations and multiple outcome heads.

Critical weaknesses
  1. The subspace decomposition lacks identifiability guarantees, as neural networks may encode the same information in both branches through nonlinear transformations while maintaining low linear correlation. The paper lacks theoretical and empirical validation of this issue, adding theoretical analysis and nonlinear dependence metrics such as HSIC and mutual information may be clearer.

  2. Although the paper claims that source-domain data improve estimation when target-domain samples are limited, the proposed method appears to use source data mainly to estimate μ0\mu_0 and μ1\mu_1, rather than directly reducing the sampling error of the empirical target-domain distribution.

  3. The paper lacks theoretical analysis of how group-wise domain discrepancy affects Target-Domain ATE estimation error and how outcome prediction error, domain discrepancy, and representation loss are related, leaving the method's rationale largely supported by intuition and simulations.

  4. The paper lacks experiments on semi-synthetic and real-world datasets to validate the effectiveness and generalizability of the proposed method.

  5. The paper includes too few baselines, particularly DeR-CFR, which also employs representation decomposition, and other classical and state-of-the-art methods.

Questions for authors

Please refer to weakness.

2 comments

Thank you for identifying where the interpretation and evaluation of our method need to be strengthened.

  1. We agree that low linear correlation does not rule out nonlinear redundancy. We use the orthogonality penalty to discourage linear overlap between ZZ and EE within each domain. Since both branches feed the outcome heads, the penalty is intended to encourage them to contribute complementary features. It does not guarantee that the branches encode distinct information or identify the causal role of either branch. We will revise the manuscript to describe the branches by how they enter the prediction and alignment objectives. In future work, we will use HSIC on held-out data to assess nonlinear dependence between the two branches and report it alongside target-domain treatment effect estimation errors.
  2. Source observations contribute to fitting the two treatment-specific conditional outcome means. We then average the fitted effects over target covariates only. For a fixed target sample, incorporating source data does not change the sampling error of the outer average of the true conditional effects. Our simulation metric compares the fitted and oracle effects averaged over the same target test units. We have corrected the manuscript to distinguish this metric from the error relative to the target-population ATE.
  3. For a fitted representation, under treatment overlap, within-arm source coverage, and bounded loss, the bound separates source prediction error, the within-arm discrepancy in the aligned representation, residual differences in prediction risk given that representation, and information lost by the full representation. Alignment targets the discrepancy term. Reducing it tightens the bound if the other terms remain unchanged. We thank the reviewer for this suggestion. In future work, we plan to formalize this decomposition in a finite-sample error bound and investigate sufficient conditions for consistency of the target ATE estimator.
  4. Thank you for this suggestion. We have added IHDP-Select to the revised Model Evaluation section. For each of the 100 IHDP benchmark realizations, we form disjoint source and target candidate pools and sample 300 source records and 100 target records before splitting the target sample for training and testing. To induce covariate shift, we draw 75% of the source records from above the median of the first covariate in the source pool. The target sample is stratified by observed treatment and does not use this quota. We retain the original covariates, treatment assignments, and benchmark outcomes and potential-outcome means. Thus, the shift comes from sample selection while the outcome-generating mechanism remains fixed. GADANet achieves mean ATE errors of 0.2829 with TARNet and 0.3263 with DragonNet, reductions of 68.8% and 76.7% relative to target-only training. It also lowers mean error on all four reported metrics relative to Warm-start with both backbones. Compared with the global-alignment ablation, mean ATE error falls by 29.4% and 20.9%, respectively. These results extend our evidence beyond fully synthetic covariates and support the use of treatment-conditional alignment when labeled target data are scarce.
  5. Thank you for this suggestion. In Section 4 of the revised manuscript, we added Counterfactual Regression (CFR), which balances treated and control representations within the target domain. GADANet outperforms CFR in treatment effect estimation on both the synthetic datasets and the semi-synthetic IHDP-Select benchmark. This adds an established causal representation learning baseline to our original comparisons with target-only training, pooling, and warm-start. We agree that DeR-CFR is a relevant comparator because it also decomposes the learned representation, and we plan to compare GADANet directly with DeR-CFR under the same evaluation protocol in future work.

Thank you for the detailed response and the additional IHDP-Select and CFR experiments, which meaningfully strengthen the empirical evaluation. However, several central concerns remain only partially addressed, particularly the lack of nonlinear dependence analysis, a formal finite-sample bound, real-world validation, and direct comparison with DeR-CFR.

Reviewer_TWRaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses average treatment-effect estimation when a small target dataset can be supplemented with a larger source dataset whose covariate distribution differs. The central challenge is to reduce domain discrepancies while retaining information needed for target-population effect estimation. The paper proposed GADANet, a framework that divides the representation into alignment and treatment-effect branches, with an orthogonality penalty encouraging separation. Two domain discriminators align source and target representations separately within the treated and control groups, while shared outcome heads use both branches. This design aims to avoid cross-treatment alignment and reduce negative transfer. Experiments integrate the framework with TARNet and DragonNet across eight configurations of a synthetic generator, varying target sample size and nuisance-only or full covariate shift. Comparisons with target-only training, pooling, and warm-start transfer report improved ATE estimation in many settings. Ablations assess the contributions of group-specific alignment and orthogonality, with results that vary across configurations and backbones.

Critical strengths
  1. The paper considers a meaningful and interesting problem. Estimating treatment effects with limited target-domain data is an important setting for causal transfer learning. A larger source dataset can help, but differences in covariate distributions can also make transfer unreliable. The paper addresses the relevant challenge of using source information while preserving what is needed to estimate the target effect. Distinguishing nuisance-only shifts from shifts that change the target ATE further clarifies the problem being studied.

  2. The method is overall reasonable, and the experiments are overall supportive. Aligning domains separately within treatment groups is a sensible way to avoid mixing treated and control observations. Keeping an effect branch outside direct adversarial alignment also provides a reasonable way to retain predictive information. The evaluation uses two backbones, eight simulation configurations, and 50 runs per configuration. The decrease on ATE error for baselines such as TARNet and DragonNet relative to target-only training overall supports the method's usefulness within the tested settings.

Critical weaknesses
  1. The paper provides no theoretical guarantee for target-domain generalization or preservation of treatment-effect information. Section 3 describes the architecture and objective, but gives no error bound, consistency result, or sufficient conditions under which alignment improves target ATE estimation. Alignment and orthogonality alone do not ensure informative features, since constant representations can satisfy both. Although the outcome loss discourages complete collapse, the objective does not establish that the two branches retain the information required for causal estimation or that transfer avoids increasing target error. The experiments provide support for the tested simulations, but leave unclear which properties of a new source-target pair would make the method reliable. This limits the justification for generalizing beyond the reported settings.

  2. The architectural novelty requires further clarification and the comparisons do not fully showcase its advantage over baselines. Domain Separation Networks already combine separate representations with orthogonality, and Multi-Adversarial Domain Adaptation uses group-specific discriminators. GADANet adapts these ideas to observed treatment groups and causal outcome heads. However, the main comparators are target-only training, pooling, and warm-start transfer. They do not isolate the benefit of the proposed combination relative to the closest conditional-alignment designs. The empirical gains therefore provide limited evidence for a distinct methodological advance.

Questions for authors
  1. In Equation (2), how is domain alignment balanced against keeping the representation branches distinct and retaining treatment-effect information? Constant features can satisfy alignment and orthogonality. Beyond combining losses and matching their magnitudes in Appendix B.1, further discussions are suggested on to which extent these targets harm representation informativeness.

  2. Under the stated covariate-shift assumptions, what justifies expecting lower target ATE error after alignment? Could authors estiablish some theoretical results to guarantee the consistency or the target-domain generalization under certain additional assumptions?

  3. How were the Table 9 adversarial strengths and balancing schemes selected? Were the baselines and ablations given comparable tuning budgets, and were the ablations retuned independently?

2 comments

We thank the reviewer for raising these questions.

  1. We agree that constant features can satisfy both domain alignment and decorrelation. Our design leaves EE available to the outcome heads without directly aligning its distribution. Exact alignment of the full representation (Z,E)(Z,E) within a treatment group would also force the predicted effects h1(Z,E)−h0(Z,E)h_1(Z,E)-h_0(Z,E) to have the same distribution across domains within that group. This can conflict with accurate prediction in both domains when shifted effect modifiers change the distribution of true conditional effects. Partial alignment leaves the outcome heads free to use those predictors through EE. Factual supervision encourages the joint representation to retain predictive information, but one branch may still learn a nearly constant representation if the other already supports accurate outcome prediction. Strong alignment or decorrelation can also suppress useful predictive signals. We acknowledge that the calibration schemes in Appendix B.1 adjust the relative scales of the objectives; they do not guarantee preservation of treatment-effect information.
  2. Under target treatment overlap and a representation sufficient for the shared conditional outcome means, a natural approach is to bound target prediction error using three terms: source conditional-mean estimation error, the within-treatment discrepancy in ZZ, and residual differences in prediction risk conditional on ZZ. Grouped alignment targets the second term. Reducing that discrepancy can tighten the bound when the other terms are controlled, but it does not guarantee lower error for a particular fitted model. In future work, we plan to formalize this argument with a finite-sample error bound and establish sufficient conditions for consistency by identifying when both the conditional-mean estimation error and the target-sample averaging error vanish. 3.For each synthetic setting in Table 9, we selected the GRL strength λ\lambda and the balancing scheme through manual trials of the full model. For settings with larger covariate shifts between the source and target domains, we extended the search range for λ\lambda to include higher values, allowing cross-validation to evaluate stronger alignment.Appendix B.1 describes how the selected scheme determines β\beta. The auto_beta scheme updates β\beta during training using an exponential moving average of mini-batch loss ratios. The gradient_norm scheme computes β\beta before training from the ratio of average gradient norms of the outcome and domain losses on the alignment branch, then holds it fixed for that run. The synthetic experiments used a common base learning rate of 10−310^{-3} and a 300-epoch training limit. We did not use a matched hyperparameter search budget across methods. The ablations inherited the applicable settings from the full model and were not retuned independently. The ablation results therefore compare the variants at the settings selected for the full model.
  3. We agree that representation splitting, orthogonality penalties, and multiple discriminators are established techniques. Our contribution is a formulation of partial, treatment-conditional alignment for causal transfer with few labeled target observations. GADANet aligns only ZZ within observed treatment groups, while both outcome heads use (Z,E)(Z,E). DSN uses only its shared branch for task prediction, whereas MADA weights its domain discriminators using predicted class probabilities. We have clarified these distinctions in the revised Related Work section. DSN and MADA were developed for classification without labeled target data during training. GADANet instead uses observed treatments and factual outcomes from both domains to estimate treatment effects. A controlled comparison would therefore require adapting DSN and MADA to the same outcome prediction task and supervision. Target-only training, pooling, and warm-start assess the benefit of transfer, while CFR provides a within-domain causal comparator. GADANet-GA offers a more direct test of grouping: on IHDP-Select, GADANet reduces mean ATE error relative to global alignment by 29.4% with TARNet and 20.9% with DragonNet. This comparison supports treatment-conditional alignment in that setting. We agree that it does not isolate the benefit of aligning only one branch. A closer control would keep the same treatment groups and outcome heads while applying alignment to the full representation (Z,E)(Z,E).

Thank you for the additional explanations and experimental results. I appreciate the clearer discussion of the method’s scope and limitations. However, as the authors also acknowledge, key concerns remain unresolved. The objective does not guarantee preservation of treatment-effect information or improved target-domain ATE estimation, and the proposed theoretical justification remains future work. The lack of matched tuning budgets and independently retuned ablations also limits conclusions about the components’ benefits. Finally, the added global-alignment comparison supports treatment-specific grouping but does not isolate the advantage of partial alignment or establish its benefit over closely related approaches. These concerns hinder the method’s justification and distinct contribution.

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

This paper studies average treatment effect estimation with limited target-domain data and proposes GADANet. The method separates representations into domain-alignment and treatment-effect branches, uses an orthogonality penalty to reduce their correlation, and performs adversarial domain alignment separately within the treated and control groups. Two outcome heads use the full representation for prediction. Across eight simulated configurations with TARNet and DragonNet, the authors report improvements over target-only training, pooling, and warm-start transfer.

Critical strengths
  1. The paper focuses on covariate shift, providing a reasonably clear scope.

  2. Leveraging external data under target-domain scarcity while avoiding negative transfer is a meaningful problem.

Critical weaknesses
  1. In Sections 3.2–3.3, the penalty constrains second-order correlations and the outcome loss uses both branches without specifying where information should reside. What is the relationship between this and causality? Second-order correlation and causality are clearly not equivalent. At the same time, orthogonal regularization cannot determine which of the two branches carries interference information and which carries effect information, respectively.

  2. The authors match Ps(Z∣T=g)P_s(Z\mid T=g) and Pt(Z∣T=g)P_t(Z\mid T=g). For this objective, conditioning on treatment restricts which distributions are aligned; it does not ensure that the removed differences are irrelevant to treatment effects.

  3. Algorithm 1 ultimately averages only the predicted treatment effects for the target test samples. Although source-domain data helps in learning the conditional outcome function, this is not the same as reducing sampling error in the distribution of target covariates.

  4. The contribution beyond existing domain separation and conditional adversarial learning remains unclear. The paper already cites Domain Separation Networks, so representation separation and orthogonality constraints should not themselves be presented as new mechanisms. Group-specific discriminators also require comparison with approaches such as Multi-Adversarial Domain Adaptation.

  5. Table 5 claims that the full method outperforms all three baselines with both backbones, although Pooling has lower TARNet ATE error than the full method, 0.2359 versus 0.2531.

Questions for authors
  1. How does the second-order decorrelation penalty establish the claimed causal roles of the two branches? What ensures that one captures nuisance information and the other retains treatment-effect information, given that the outcome heads use both?

  2. Why should matching Ps(Z∣T=g)P_s(Z\mid T=g) and Pt(Z∣T=g)P_t(Z\mid T=g) remove only treatment-effect-irrelevant differences? Under what conditions does this alignment preserve relevant information when effect modifiers also shift across domains?

  3. How does the proposed estimator reduce the target-distribution sampling error discussed in Section 3.1, given that Algorithm 1 averages predictions only over target test samples? Please distinguish this from improving conditional outcome estimation.

  4. What substantive methodological contribution does GADANet provide beyond combining domain separation with treatment-conditional adversarial alignment? How does it differ from Domain Separation Networks and Multi-Adversarial Domain Adaptation?

  5. Could the authors reconcile the claim in Table 5 with Pooling’s lower TARNet ATE error, 0.2359 versus 0.2531 for GADANet, and clarify the scope of the claimed improvements?

1 comment

Thank you for identifying where our claims went beyond what the objective and experiments support. We address the five points below.

  1. Interpretation of the two branches. We agree that low cross-branch correlation does not rule out nonlinear redundancy. The orthogonality penalty discourages empirical linear correlations; it does not establish independence or identify which branch contains effect-modifying information. The distinction between the branches is in how they are trained: the domain discriminators see only (Z), while both outcome heads use ((Z,E)). We will revise Sections 3.2–3.3 to describe the branches by their training roles and remove the claim that the penalty separates information by its causal role. 2.Alignment when effect modifiers shift. Matching Ps(Z∣T=g)P_s(Z\mid T=g) and Pt(Z∣T=g)P_t(Z\mid T=g) compares the domains within the same observed treatment group. It does not distinguish differences that matter for treatment effects from those that do not. GADANet applies adversarial alignment only to ZZ, while both outcome heads use the full representation (Z,E)(Z,E). A sufficient condition is that ZZ can be aligned within each treatment group while both conditional potential-outcome means remain expressible as Ed[Y(g)∣X]=hg(Z,E)\mathbb E_d[Y(g)\mid X]=h_g(Z,E), with the same hgh_g across domains for each a∈{0,1}a\in\{0,1\}. A shifted effect modifier can then still inform prediction through EE, which is not directly aligned. This condition shows when alignment and outcome prediction are compatible. We acknowledge that GADANet does not guarantee that the learned representation preserves the information needed to estimate treatment effects when effect modifiers shift across domains.
  2. Source data and the outer ATE average. The reviewer is correct. Let (\tau(x)=\mu_1(x)-\mu_0(x)) and let (\mathbb P_{t,n}) denote the empirical distribution of the target covariates used for evaluation. The population ATE error decomposes as [ \mathbb P_{t,n}\widehat\tau-P_t\tau =\mathbb P_{t,n}(\widehat\tau-\tau)+(\mathbb P_{t,n}-P_t)\tau. ] Source data can improve estimation of the conditional outcome functions and thereby affect the first term. It does not change the second term for a fixed target sample. We have already revised Section 3.1 to distinguish this metric from the error relative to the target-population ATE.
  3. Relationship to DSN and MADA. We agree that representation splitting, decorrelation, and multiple domain discriminators are established techniques. Our contribution is a transfer learning framework for estimating target-domain treatment effects when labeled target data are scarce. We uses observed outcomes from both domains to fit the two treatment-specific conditional outcome functions and applies adversarial alignment to only part of the representation to mitigate the risk of negative transfer under covariate shift. DSN (Bousmalis et al., 2016) and MADA (Pei et al., 2018) were formulated for unsupervised domain adaptation in image classification. They use labeled source data and unlabeled target inputs during training, with target class prediction as the objective. In our setting, treatment and factual outcomes are observed in both domains, but the target sample is small. We estimate the conditional outcome mean under each treatment and average their fitted differences over target covariates to estimate the target ATE. These differences also shape the training objective. DSN uses the shared branch for classification and both branches for reconstruction. In GADANet, both branches contribute to each outcome prediction. MADA weights discriminator inputs by predicted class probabilities, whereas GADANet uses observed treatment to route samples and aligns only (Z) within each treatment arm. We have revised the Related Work section to compare GADANet with these methods and clarify our contribution: a formulation of partial, treatment-conditional domain alignment for causal transfer.
  4. Correction to Table 5. We will correct the caption and qualify the claims by backbone, metric, and comparison method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.