acceptodds
Under review as a conference paper at ICLR 2027

MultiCFR: Counterfactual Regression for Multimodal Treatment Effect Estimation

Abstract

Existing individual treatment effect (ITE) estimation methods mainly focus on tabular covariates. However, in practice, tabular data often coexist with text and image modalities, particularly in healthcare. Leveraging such multimodal data can capture richer confounding information and enable more effective confounding adjustment. Only a few methods transfer multimodal fusion techniques to ITE estimation, but overlook explicit tabular semantics and the trade-off between distribution balancing and factual outcome prediction. To address these issues, we propose , a counterfactual regression method for multimodal ITE estimation with two key designs. (1) Tabular semantic anchoring uses explicit tabular semantics to guide relevant information extraction from text tokens and image patches. (2) Dual-branch representation learning separates the fused representation into balanced and predictive branches to mitigate confounding bias while preserving useful information for factual outcome prediction. Theoretically, we derive a generalization-error bound that characterizes the trade-off between distribution balancing and predictive information preservation. Empirically, we construct two semi-synthetic multimodal datasets for standardized evaluation, MIMIC-TIT and IHDP-TIT, where MultiCFR consistently outperforms existing baselines with lower ITE estimation errors.

Primary area:
causal reasoning
Keywords:
Individual treatment effect, counterfactual regression, causal representation learning, multimodal causal inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CPnMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors proposed MultiCFR to estimate treatment effects from tabular, textual, and image covariate. They used a patient-specific tabular anchor to guide the attention over text tokens and image patches. Then they divided the fused representation through a soft gate into a balance-regularized branch and an unregularized predictive branch. They also presented an error-bound decomposition and evaluate their proposed method using two semi-synthetic benchmarks, MIMIC-TIT and IHDP-TIT.

Critical strengths

The main experimental results show improvements over several baselines. The evaluation also includes useful ablation studies on anchors, balancing, and modalities which is very nice to demonstrate the performance of the proposed method.

Critical weaknesses
  1. The main novelty of this proposed method lies in combining tabular-conditioned attention with selective representation balancing. However, the trade-off between balancing and prediction is already well established, including in DIGNet, which the paper cites, and the novelty of the two core designs is not sufficiently demonstrated. Direct comparisons are needed with standard tabular conditional attention and disentangled representation methods such as DIGNet, using the same encoder and parameter budget. The current results do not clearly show whether the improvements come from the semantic anchors, general conditional retrieval, or a stronger fusion network.

  2. Complementary soft masks and separate projections do not guarantee that the branches learn distinct information. The balanced branch may become nearly constant, while the unbalanced branch captures most predictive information. Moreover, the residual term ΔaB\Delta_a^B is not directly controlled during training. This reveals a potential degeneracy and a gap in the theoretical guarantee, but does not establish that collapse occurs in practice. Additionally, some important methods such as DISCRET have been cited but not compared. TransTEE (Zhang, Yi-Fan, et al. "Exploring transformer backbones for heterogeneous treatment effect estimation." arXiv preprint arXiv:2202.01336 (2022).) is another missing baseline model.

Questions for authors
  1. Do the two branches learn distinct information? Can you provide diagnostics for branch collapse, treatment prediction, information retention, and residual differences?

  2. How does MultiCFR compare with standard covariate-conditioned attention and disentangled representation methods under the same encoders and parameter budgets?

  3. Can you report results from the full-tabular oracle and test robustness to different confounding strengths, modality-specific information, and noise levels? Additionally, reporting uncertainty estimates for the main ablations and provide more details on implementation and baseline adaptations would be also helpful?

5 comments
AuthorsOpenReview ↗

Response [1 / 2]

We sincerely thank you for your constructive feedback, which has helped us conduct a more comprehensive evaluation. Below are our point-by-point responses.

Q1&W2: Clarification on the Dual-Branch Design

R1: We clarify that MultiCFR employs soft allocation rather than causal disentanglement; thus, the branches need not contain mutually exclusive information. This avoids discarding overlapping confounding and prognostic signals, without suppressing predictive information.

Following your valuable suggestion, we provide collapse, treatment prediction, information retention, and residual diagnostics. The experimental results show that: (1) No branch collapse (Tab. 1): Both branches maintain nonzero feature variance and effective rank, with only 0.67% saturated gates. (2) Treatment prediction (Tab. 2): ZBZ^B and ZRZ^R achieve nearly identical treatment AUCs (0.691/0.690), confirming that treatment information is not forcibly separated. (3) Information retention (Tab. 3): The joint representation achieves the highest prognostic-score and true-ITE R2R^2 (0.880/0.843), supporting the benefit of soft allocation. (4) Residual discrepancy (Tab. 4): MultiCFR reduces the weighted residual discrepancy from 0.177 to 0.162 (8.5%) and ATE error by 28.1% relative to global balancing, demonstrating effective selective balancing despite information overlap.

Tab. 1 Branch Collapse

ZBZ^BZRZ^R
Feature Variance0.0520.080
Effective Rank2.852.70

Tab. 2 Treatment Prediction

Treatment AUC
ZBZ^B0.691
ZRZ^R0.690

Tab. 3 Information Retention (R2R^2)

ZBZ^BZRZ^RJoint
Prognostic Score0.8750.8720.880
True ITE0.8360.8340.843

Tab. 4 Residual Discrepancy

Weighted Residual
No Balance0.172±0.015
Global Balance0.177±0.014
MultiCFR0.162±0.011

Q2&W1: Comparison with Standard Attention and Disentangled Representation Methods

R2: Comparison with standard attention. We conducted a controlled 2×2 comparison (Tab. 5), varying only the attention mechanism and balancing strategy while keeping the remaining architecture and training settings consistent. (1) Semantic anchoring: Unlike standard covariate-conditioned attention, our semantic anchor uses explicit tabular semantics to guide patient-specific retrieval of relevant text and image features. Under global balancing, it reduces ϵPEHE\sqrt{\epsilon_{PEHE}} by 6.5% and ATE error by 35.8%, validating semantic-guided retrieval. (2) Selective balancing: By balancing only ZBZ^B while preserving predictive information, it further reduces ϵPEHE\sqrt{\epsilon_{PEHE}} by 2.3%, ATE error by 28.1%, and factual RMSE by 1.6% with semantic anchoring fixed. Together, both designs achieve the best results across all three metrics, demonstrating their complementary benefits beyond a stronger fusion backbone.

Tab. 5 Comparison of Two Key Designs

Global BalancingSelective Balancing
Standard Conditional Attention0.370/0.260/1.2340.356/0.128/1.239
Semantic Anchor0.346/0.167/1.2170.338/0.120/1.198

Values denote ϵPEHE\sqrt{\epsilon_{PEHE}} / ϵATE\epsilon_{\mathrm{ATE}} / Factual RMSE.

Comparison with disentangled representations. Under a shared multimodal encoder and comparable parameter budgets, we additionally compared MultiCFR with DR-CFR, DeR-CFR, and DIGNet. As shown in Tab. 6, MultiCFR achieves the lowest ATE error (0.120) and factual RMSE (1.198). Notably, it reduces PEHE by 14.6% and ATE error by 50.8% compared with DIGNet. These results highlight the effectiveness of our anchor-guided soft allocation and selective balancing in preserving predictive information while improving treatment-effect estimation.

Tab. 6 Comparison with Disentangled Representation Methods

MethodϵPEHE\sqrt{\epsilon_{PEHE}}ϵATE\epsilon_{\mathrm{ATE}}Factual RMSE
DR-CFR0.332±0.0510.200±0.0851.228±0.005
DeR-CFR0.345±0.0530.207±0.0741.229±0.007
DIGNet0.396±0.0410.244±0.1081.221±0.005
MultiCFR0.338±0.0320.120±0.0211.198±0.008
AuthorsOpenReview ↗

Response [2 / 2]

Q3: Oracle, Robustness, and Uncertainty Evaluation

R3: We added the following additional experiments (Tabs. 7–11): (1) Full-tabular oracle. MultiCFR reduces PEHE by 18.4% compared with the 9-variable tabular baseline, demonstrating its ability to recover useful information from text and images. The Full-25 oracle serves as a full-information reference: its substantially lower error (1.254 vs. 3.121) quantifies the remaining gap between recovering information from transformed multimodal views and directly observing the original covariates.

Tab. 7 Full-tabular Oracle Results

ϵPEHE\sqrt{\epsilon_{PEHE}}ϵATE\epsilon_{\mathrm{ATE}}Factual RMSE
Tabular-9 CFR3.824±0.6160.505±0.0833.965±0.570
MultiCFR3.121±0.5450.430±0.0773.253±0.483
Full-25 CFR1.254±0.1730.254±0.0371.681±0.149

(2) Confounding strength. MultiCFR consistently outperforms DoubleML-Deep and DIGNet across all confounding strengths (α∈0,0.5,1,2\alpha\in{0,0.5,1,2}), maintaining the lowest PEHE even under strong confounding and demonstrating robust treatment effect estimation.

Tab.8 Confounding Strength Robustness (ϵPEHE\sqrt{\epsilon_{PEHE}})

Methodα=0\alpha=0α=0.5\alpha=0.5α=1\alpha=1α=2\alpha=2
DoubleML-Deep5.710±0.8935.708±0.8915.751±0.8945.783±0.893
DIGNet3.074±0.5213.083±0.4823.357±0.5543.659±0.573
MultiCFR3.033±0.5143.052±0.5053.240±0.5543.626±0.595

(3) Modality-specific confounding. With α=1\alpha=1, we vary the modality driving treatment assignment. MultiCFR consistently achieves the lowest PEHE, outperforming DoubleML-Deep and DIGNet across all settings, demonstrating robustness to modality-specific confounding.

Tab. 9 Robustness to modality-specific information Assignment (ϵPEHE\sqrt{\epsilon_{PEHE}})

Confounding-information sourceDoubleML-DeepDIGNetMultiCFR
Tabular5.747±0.8903.157±0.5173.100±0.513
Text5.730±0.8923.186±0.5243.125±0.537
Image5.730±0.8943.198±0.5493.105±0.525
All modalities5.751±0.8943.357±0.5543.240±0.554

(4) Noise robustness. MultiCFR achieves the lowest ϵPEHE\sqrt{\epsilon_{PEHE}} at all nonzero noise levels. At σ=1\sigma=1, its ϵPEHE\sqrt{\epsilon_{PEHE}} increases by 22.81% relative to the clean setting, compared with 30.25% for DIGNet. MultiCFR also obtains the lowest ϵPEHE\sqrt{\epsilon_{PEHE}}–noise AUC (3.489), compared with DIGNet (3.518) and DoubleML-Deep (5.710), indicating stronger overall ITE-estimation performance across the noise interval. Although DoubleML-Deep exhibits almost no relative degradation, its consistently high ϵPEHE\sqrt{\epsilon_{PEHE}} shows that low sensitivity alone does not imply accurate estimation.

Tab. 10 Robustness to Input Noise

Methodσ=0\sigma=0σ=0.5\sigma=0.5σ=1\sigma=1R(1)R(1)ϵPEHE\sqrt{\epsilon_{PEHE}}–noise AUC
DoubleML-Deep5.712±0.8895.709±0.8885.708±0.888−0.06%5.710±0.888
DIGNet3.009±0.4863.574±0.5683.919±0.61530.25%3.518±0.562
MultiCFR3.121±0.5453.524±0.5683.833±0.61022.81%3.489±0.568

(5) Ablation uncertainty. We report paired 95% confidence intervals for ϵPEHE\sqrt{\epsilon_{PEHE}} differences. Removing text or images consistently increases ϵPEHE\sqrt{\epsilon_{PEHE}}, providing statistical evidence for the contribution of multimodal information.

Tab. 11 Ablation Effects with Paired 95% CIs

AblationΔϵPEHE\Delta\sqrt{\epsilon_{PEHE}}95% CI
w/o Text+0.400[0.139,0.662]
w/o Image+0.470[0.270,0.670]
Tabular Only+0.703[0.405,1.001]
w/o Anchor+0.035[-0.137,0.206]

(6) Implementation details. MultiCFR uses frozen BioClinicalBERT and ResNet-18 encoders with cached text and image features, and is trained with AdamW using validation factual MSE for model selection. For fair comparison, all methods use identical multimodal inputs, preprocessing, data splits, and evaluation metrics. When adapting baselines to multimodal ITE estimation, we replace only their input modules while preserving their original mechanisms, including nuisance estimation, representation balancing or disentanglement, propensity modeling, and outcome prediction. We will provide additional implementation and baseline-adaptation details in the revision.

W2: Comparison with DISCRET and TransTEE

R4: We compare MultiCFR with DISCRET and TransTEE, and MultiCFR maintains strong performance.

Tab.12 Comparison with Additional Baselines on MIMIC-TIT

MethodϵPEHE\sqrt{\epsilon_{PEHE}}ϵATE\epsilon_{\mathrm{ATE}}Factual RMSE
TransTEE0.443±0.0910.376±0.1001.266±0.005
DISCRET0.817±0.0580.738±0.0541.748±0.006
MultiCFR0.338±0.0320.120±0.0211.198±0.008

Dear Reviewer CPnM,

We greatly appreciate your time and effort in reviewing our work. We hope that our responses have clarified our work and addressed your concerns. If so, we would sincerely appreciate your consideration of updating your rating accordingly. If you have any further concerns, we would be happy to provide further clarification.

We highly value every exchange during the review process and sincerely hope to receive further feedback from you at your earliest convenience.

Sincerely,

The Authors

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

This paper studies individual treatment effect estimation from multimodal observational data comprising tabular, text, and image inputs. The proposed MultiCFR uses tabular embeddings to retrieve relevant text and image features through cross-attention. A soft gate then splits the fused representation into an MMD-regularized balancing branch and an unregularized predictive branch. The authors derive a PEHE generalization bound under selective balancing and evaluate the method on two newly constructed semi-synthetic benchmarks, MIMIC-TIT and IHDP-TIT.

Critical strengths
  1. The problem is practically relevant, particularly for clinical settings combining structured records, notes, and medical images.

  2. MIMIC-TIT and IHDP-TIT are useful benchmark contributions, with detailed documentation for future multimodal causal inference research.

  3. The paper is clearly written and includes extensive comparisons, ablations, and visualizations showing strong performance on the proposed benchmarks.

Critical weaknesses
  1. Limited novelty in tabular semantic anchoring. The mechanism is standard cross-attention with residual gating: tabular features provide queries, while text and image tokens provide keys and values. This is a reasonable multimodal design but does not explicitly account for treatment assignment or confounding. Proposition C.1 likewise follows directly from the Softmax score-margin property and provides little additional theoretical insight.

  2. Selective balancing permits confounding leakage. Representation splitting has already been explored in causal inference. MultiCFR instead uses element-wise gating, $ z_i^B=r_i\odot s_i,\qquad z_i^R=r_i\odot(1-s_i), $ without explicit disentanglement constraints. Since zRz^R feeds the outcome predictors but receives no MMD penalty, the model can route confounding features into this branch to reduce factual loss while making zBz^B trivial. The objective does not rule out this shortcut.

  3. The theory neither establishes a tighter bound nor captures multimodal structure. Theorem 2 extends Shalit et al. (2017) through an intermediate conditional expectation and the triangle inequality. It replaces the global discrepancy with an MMD term dBd_B and an unobservable, unregularized residual shift ΔaB\Delta_a^B. Routing information into ZRZ^R may reduce dBd_B while increasing ΔaB\Delta_a^B, so a tighter overall PEHE bound is not established. Moreover, Lemma 1 and Theorem 2 treat the input as an abstract variable and apply equally to unimodal data, without accounting for cross-modal redundancy, noise, or complementary confounding.

Questions for authors
  1. What prevents confounding information from being routed into the unregularized branch zRz^R? Please report treatment-prediction AUC separately for zBz^B and zRz^R.

  2. Under what formal conditions does selective balancing yield a strictly tighter bound than global balancing, given that reducing dBd_B may increase the uncontrolled residual ΔaB\Delta_a^B?

  3. How does MultiCFR compare with DR-CFR or DeR-CFR using the same multimodal encoder backbone, both conceptually and empirically?

  4. How robust is the method when the tabular anchor is noisy, highly incomplete, or contains few features, given that both cross-modal retrieval and branch allocation depend on it?

4 comments

We sincerely appreciate your effort in providing these valuable suggestions. We have added further theoretical analysis and experimental results, and hope the following clarifications will help resolve your concerns.

Q1&W2: Clarification of the Dual-Branch Information Allocation

R1: We clarify that MultiCFR employs soft allocation rather than causal disentanglement. Thus, confounding information is not explicitly prohibited from entering ZRZ^R. Instead, balancing is applied selectively to ZBZ^B, while ZRZ^R retains complementary predictive information, avoiding excessive suppression of outcome-relevant information.

Following your suggestion, we evaluate treatment-prediction AUC for each branch. The similar AUCs indicate that treatment-related information is retained in both branches rather than strictly separated. Nevertheless, MultiCFR reduces residual discrepancy by 8.5% and ATE error by 28.1% compared with global balancing, demonstrating the empirical benefits of selective balancing despite overlapping information.

Tab. 1 Treatment Prediction and Residual Discrepancy

EvaluationSettingResult
Treatment PredictionZBZ^B0.691
Treatment PredictionZRZ^R0.690
Residual DiscrepancyNo Balance0.172±0.0150.172\pm0.015
Residual DiscrepancyGlobal Balance0.177±0.0140.177\pm0.014
Residual DiscrepancyMultiCFR0.162±0.0110.162\pm0.011

Q2&W3: Conditions for Tighter Selective-Balancing Bounds

R2: Below, we provide conditions for tighter selective-balancing bounds.

Theoretical Analysis

Formal comparison. Let j∈{S,G}j\in\{S,G\} denote selective and global balancing. Define Fj=F0j+F1jF^j=F_0^j+F_1^j, Kj=πB0j+(1−π)B1jK^j=\pi B_0^j+(1-\pi)B_1^j, and Dj=πΔ0j+(1−π)Δ1jD^j=\pi\Delta_0^j+(1-\pi)\Delta_1^j. By Theorem 2, the corresponding upper bounds are Uj=2(Fj+Kjdj+Dj)U_j=2(F^j+K^jd_j+D^j). Thus, US<UGU_S<U_G if and only if (FG−FS)+(KGdG−KSdS)>DS−DG.(F^G-F^S)+(K^Gd_G-K^Sd_S)>D^S-D^G.

Sufficient condition. Assume ∣ℓaCF,S(w)−ℓaF,S(w)∣≤δa|\ell_a^{CF,S}(w)-\ell_a^{F,S}(w)|\leq\delta_a for q1−aB,Sq_{1-a}^{B,S}-almost every ww. Then ΔaS≤δa\Delta_a^S\leq\delta_a, and a sufficient condition for US<UGU_S<U_G is (FG−FS)+(KGdG−KSdS)>πδ0+(1−π)δ1.(F^G-F^S)+(K^Gd_G-K^Sd_S)>\pi\delta_0+(1-\pi)\delta_1.

This ensures that improvements in factual prediction and distribution balance outweigh the bounded residual discrepancy, without requiring ZRZ^R to be confounder-free.

Empirical Support

Compared with global balancing, MultiCFR reduces ϵPEHE\sqrt{\epsilon_{PEHE}} (0.346 →\rightarrow 0.338), ATE error (0.167 →\rightarrow 0.120), factual RMSE (1.217 →\rightarrow 1.198), and estimated residual discrepancy (0.177 →\rightarrow 0.162). These consistent improvements demonstrate the empirical advantage of selective balancing over global balancing.

Q3: Comparison with DR-CFR and DeR-CFR

R3: Conceptually, DR-CFR and DeR-CFR decompose representations according to distinct causal roles (instrumental, confounding, and adjustment factors). However, MultiCFR employs soft information allocation rather than explicit causal-role decomposition, selectively balancing ZBZ^B while preserving outcome-relevant information in ZRZ^R to avoid excessive information suppression.

Empirically, using the same multimodal encoder backbone, MultiCFR reduces ATE error by 40.0% and 42.0% over DR-CFR and DeR-CFR, respectively, while achieving the lowest factual RMSE (1.198) and competitive PEHE.

Tab. 2 Comparison with DR-CFR and DeR-CFR

MethodϵPEHE\sqrt{\epsilon_{PEHE}}ϵATE\epsilon_{\mathrm{ATE}}Factual RMSE
DR-CFR0.332±0.0510.200±0.0851.228±0.005
DeR-CFR0.345±0.0530.207±0.0741.229±0.007
MultiCFR0.338±0.0320.120±0.0211.198±0.008

Q4: Robustness to Tabular Anchors

R4: To assess the robustness of MultiCFR's tabular anchor, we test noise, missingness, and limited features by perturbing the anchor at test time while keeping the multimodal inputs unchanged.

The results show that MultiCFR remains stable under severe anchor incompleteness and feature reduction. Even with 90% missingness or only one retained feature, ϵPEHE\sqrt{\epsilon_{PEHE}} increases by just 5.0% and 6.2%, respectively. Under moderate Gaussian noise (σ=0.5\sigma=0.5), ϵPEHE\sqrt{\epsilon_{PEHE}} increases by 6.8%, while severe noise (σ=2\sigma=2) leads to a 41.4% increase. These results demonstrate robustness to incomplete and low-dimensional anchors.

Tab.3 Sensitivity of MultiCFR to Tabular-Anchor Degradation

Degradation typeConditionϵPEHE\sqrt{\epsilon_{PEHE}}Relative change
ReferenceClean anchor, 18 features0.338±0.032—
Measurement noiseGaussian noise, σ=0.5\sigma=0.50.361±0.062+6.8%
Measurement noiseGaussian noise, σ=2\sigma=20.478±0.094+41.4%
Incompleteness50% random missingness0.347±0.047+2.7%
Incompleteness90% random missingness0.355±0.043+5.0%
Few-feature anchor3 retained features0.356±0.042+5.3%
Few-feature anchor1 retained feature0.359±0.047+6.2%

Dear Reviewer yPEA,

We greatly appreciate your time and effort in reviewing our work. We hope that our responses have clarified our work and addressed your concerns. If so, we would sincerely appreciate your consideration of updating your rating accordingly. If you have any further concerns, we would be happy to provide further clarification.

We highly value every exchange during the review process and sincerely hope to receive further feedback from you at your earliest convenience.

Sincerely,

The Authors

Thank you to the authors for the response and the additional experiments. While I appreciate the new robustness checks (Table 3), the rebuttal results do not resolve—and in fact reinforce—my core concerns:

  1. Confounding leakage (W2 & Q1): The nearly identical treatment-prediction AUCs between the balanced and unregularized branches (0.691 vs. 0.690 in Table 1) confirm that confounding information flows freely into the unregularized branch. Labeling this as "soft allocation" does not mitigate the risk of the model bypassing the balancing penalty.
  2. Theoretical limitations (W3 & Q2): The newly derived condition for a tighter bound simply rearranges the terms of the bound definitions without offering meaningful structural insights, and the theoretical framework remains completely agnostic to multimodal structure.
  3. Comparison with causal disentanglement (Q3): When evaluated with the same multimodal backbone (Table 2), standard causal disentanglement (DR-CFR) actually outperforms MultiCFR on the primary PEHE metric (0.332 vs. 0.338), which questions the necessity and advantage of the proposed gating mechanism.

Given that these fundamental methodological and theoretical limitations remain, I will maintain my original score.

We thank the reviewer for the follow-up comments. We would like to clarify three points regarding our contributions and experimental evidence.

1. Confounding Leakage.

We acknowledge that MultiCFR does not explicitly prevent confounding information from entering ZRZ^R. However, confounding information retained in ZRZ^R does not necessarily imply confounding bias in ITE estimation. Similar treatment-prediction AUCs (0.691 vs. 0.690) show that both branches retain treatment-predictive information, but do not establish harmful confounding bias, since treatment predictability alone cannot distinguish confounding from treatment-only information.

We further clarify this distinction through both theoretical analysis and empirical evidence.

(a) Theoretical Analysis. MultiCFR predicts potential outcomes using the joint representation Z=[ZB;ZR]Z=[Z^B;Z^R]. Under consistency, conditional ignorability given XX, overlap, and outcome sufficiency, suppose E[Y(a)∣X]=ga(Z),a∈{0,1}\mathbb{E}[Y(a)\mid X]=g_a(Z),\quad a\in\{0,1\}. Then, $ \mathbb{E}[Y\mid A=a,Z]=\mathbb{E}[Y(a)\mid Z]=g_a(Z)$, implying τ(X)=g1(Z)−g0(Z)\tau(X)=g_1(Z)-g_0(Z). This argument is consistent with prior research on deconfounding scores [1] and representation-induced confounding bias [2], which emphasizes preserving causally sufficient information in learned representations. Thus, retaining confounding information in ZRZ^R does not inherently invalidate causal adjustment, provided the joint representation satisfies these conditions.

(b) Experimental Evidence. Our diagnostics do not support the extreme shortcut in which ZBZ^B becomes trivial while ZRZ^R carries all predictive information. Specifically, ZBZ^B independently retains substantial prognostic and ITE information (R2=0.875/0.836R^2=0.875/0.836), with nonzero feature variance (0.052) and effective rank (2.85). Adding ZRZ^R increases treatment-prediction AUC only marginally (0.691→0.692), while only 0.67% of gates are saturated. Furthermore, MultiCFR reduces the measured residual discrepancy by 8.5% and ATE error by 28.1% compared with global balancing. Although these results do not rule out all confounding leakage, they provide evidence against the extreme shortcut raised by the reviewer.

Tab. 1 Dual-Branch Diagnostics

MetricZBZ^BZRZ^R[ZB;ZR][Z^B;Z^R]
Treatment AUC0.6910.6900.692
Prognostic R2R^20.8750.8720.880
True ITE R2R^20.8360.8340.843
Feature Variance0.0520.080—
Effective Rank2.852.70—

[1] Clivio, O., et al. "Deconfounding Scores and Representation Learning for Causal Effect Estimation with Weak Overlap." AISTATS, 2026.

[2] Melnychuk, V., Frauen, D., and Feuerriegel, S. "Bounds on Representation-Induced Confounding Bias for Treatment Effect Estimation." ICLR, 2024.

2. Theoretical Contribution.

We clarify that MultiCFR first performs multimodal fusion, followed by causal effect estimation on the learned fused representation. Accordingly, Theorem 2 provides theoretical insight into the trade-off among factual prediction error, distribution balancing, and residual discrepancy, rather than guaranteeing the unconditional superiority of selective balancing. While this provides insight into the selective-balancing mechanism at the representation level, we acknowledge that the current theoretical framework remains agnostic to multimodal structure, which is a limitation we aim to address in future work.

3. Comparison with DR-CFR.

We acknowledge that DR-CFR achieves slightly lower root PEHE (0.332 vs. 0.338). However, MultiCFR reduces ATE error by 40.0% and factual RMSE by 2.4% compared with DR-CFR.

To further validate our design, we conducted a controlled 2×2 comparison, varying only the attention mechanism and balancing strategy. With selective balancing fixed, semantic anchoring reduces root PEHE by 5.1% and ATE error by 6.3%. Importantly, even with semantic anchoring fixed, selective balancing provides additional improvements, reducing root PEHE by 2.3% and ATE error by 28.1%. Together, the two components achieve 8.6% lower root PEHE and 53.8% lower ATE error than standard attention with global balancing, demonstrating their complementary benefits.

Tab. 2 Comparison of Two Key Designs

Attention MechanismGlobal BalancingSelective Balancing
Standard Conditional Attention0.370/0.260/1.2340.356/0.128/1.239
Semantic Anchor0.346/0.167/1.2170.338/0.120/1.198

Values denote ϵPEHE\sqrt{\epsilon_{\mathrm{PEHE}}} / ϵATE\epsilon_{\mathrm{ATE}} / Factual RMSE.

We appreciate the reviewer's careful consideration and hope these clarifications help accurately characterize the scope and contributions of our work.

Reviewer_5K5iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MultiCFR for multimodal individual treatment effect estimation using tabular, text, and image covariates. The method uses tabular semantic anchoring to guide information extraction from text and images, and a dual-branch representation to balance treatment groups while preserving outcome-predictive information. The authors evaluate MultiCFR on two semi-synthetic multimodal benchmarks, MIMIC-TIT and IHDP-TIT, comparing against five multimodal treatment-effect baselines and conducting ablation and modality analyses. MultiCFR achieves the best overall performance on both datasets, including lower PEHE and ATE error, while the ablations support contributions from both semantic anchoring and selective balancing.

Critical strengths
  1. The motivation of integrating multimodal information is very clear. The paper addresses confounding information distributed across structured and unstructured modalities.
  2. The semantic anchoring and selective-balancing ideas are intuitive and well connected to the causal objective.
  3. The evaluation is comprehensive The paper includes multiple multimodal baselines, ablations, modality analyses, and both ITE and factual prediction metrics.
Critical weaknesses
  1. The evaluation remains semi-synthetic. MIMIC-TIT uses real EHR, reports, and chest X-rays, but treatment assignments and potential outcomes are synthetically generated; IHDP-TIT also relies on constructed text and visual representations. Therefore, the results demonstrate recovery under known data-generating processes rather than effectiveness in a real clinical treatment-effect setting.
  2. The IHDP-TIT benchmark is somewhat artificial, since the text and image modalities are deterministic transformations of the original tabular covariates. It is therefore unclear how much the performance gain reflects true multimodal reasoning versus recovery of information that has been re-encoded into different modalities.
  3. The qualitative semantic-anchoring evidence is limited. Figure 2 shows that the model attends to particular X-ray regions and report tokens, but attention alone does not establish that these regions or words are causally relevant to treatment-effect estimation.
  4. The dual-branch design is intuitive, but it is unclear whether the learned split truly separates confounding-related from outcome-predictive information. Treatment-related information may remain in the predictive branch, potentially weakening the intended balancing effect. The authors might need to provide stronger evidence that the learned decomposition behaves as intended.
Questions for authors
  1. Can the authors provide stronger evidence that the semantic anchor retrieves clinically or causally relevant information rather than simply producing plausible attention patterns?
  2. How sensitive are the results to the specific multimodal confounding mechanism used to construct MIMIC-TIT? Would MultiCFR retain its advantage under substantially different treatment and outcome data-generating processes?
  3. Given that IHDP-TIT text and images are deterministic transformations of original tabular covariates, what additional challenge does this benchmark provide beyond recovering differently encoded versions of known structured information?
  4. Could the authors discuss how MultiCFR would be evaluated in a real clinical observational dataset where individual-level counterfactual outcomes are unavailable?
3 comments
AuthorsOpenReview ↗

Response [1 / 2]

We sincerely thank you for your thorough and thoughtful review, which is invaluable for improving our work. We have carefully addressed your concerns below and hope these clarifications address your questions.

Q1&W3: Validation of Semantic Anchor Relevance

R1: We added two diagnostics, focusing on factual prediction error and ITE-relevant information retention, to assess whether the semantic anchor retrieves treatment-effect-relevant information. First, in the functional intervention test (Tab. 1), retaining top-attended report tokens achieves a lower factual RMSE than random selection (Top 20% 1.224 vs. Random 20% 1.282), while removing them causes a larger error (Top 20% 1.321 vs. Random 20% 1.235). Second, as shown in Tab. 2, the DGP-information probe indicates that top-attended text is substantially more informative about true ITE than random content (R2R^2: 0.408 vs. 0.200), with a similar trend for images.

Tab. 1 Functional Intervention Test (Factual RMSE)

Text InterventionTop 20%Random 20%Bottom 20%
Retain only1.2241.2821.324
Remove1.3211.2351.224

Tab. 2 DGP-Information Enrichment

ModalitySelectionR2(μ0)R^2(\mu_0)R2(μ1)R^2(\mu_1)R2(τ)R^2(\tau)
TextTop0.5210.5190.408
TextRandom0.2710.2600.200
ImageTop0.2890.3080.266
ImageRandom0.1980.2080.171

Q2: Robustness to Confounding and Treatment/Outcome Mechanisms

R2: Confounding strength. As shown in Tab.3, MultiCFR consistently outperforms DoubleML-Deep and DIGNet across all confounding strengths (α∈0,0.5,1,2\alpha\in{0,0.5,1,2}), maintaining the lowest ϵPEHE\sqrt{\epsilon_{PEHE}} even under strong confounding and demonstrating robust treatment effect estimation.

Tab.3 Confounding Strength Robustness (ϵPEHE\sqrt{\epsilon_{PEHE}})

Methodα=0\alpha=0α=0.5\alpha=0.5α=1\alpha=1α=2\alpha=2
DoubleML-Deep5.710±0.8935.708±0.8915.751±0.8945.783±0.893
DIGNet3.074±0.5213.083±0.4823.357±0.5543.659±0.573
MultiCFR3.033±0.5143.052±0.5053.240±0.5543.626±0.595

Treatment/Outcome Mechanisms. For a fair comparison, we keep other factors unchanged, such as the treatment proportion and propensity-score dispersion, and modify only the treatment-assignment and potential-outcome generation mechanisms. Specifically, we construct a nonlinear interaction DGP with nonlinear, quadratic, and interaction effects in treatment assignment and nonlinear dependencies of potential outcomes and ITE on confounding and prognostic factors. Tab. 4 shows that MultiCFR remains competitive under this alternative DGP.

Tab.4 Comparison under an Alternative Nonlinear DGP

MethodϵPEHE\sqrt{\epsilon_{PEHE}}ϵATE\epsilon_{\mathrm{ATE}}Factual RMSE
DoubleML-Deep0.366±0.0050.185±0.0091.521±0.008
DIGNet0.360±0.0160.161±0.0301.294±0.013
MultiCFR0.348±0.0090.135±0.0161.240±0.013

Q3&W2: Clarification on the Deterministic Construction of IHDP-TIT

R3: We explain that IHDP-TIT is intentionally constructed as a controlled multimodal benchmark, following Deshpande et al. (NeurIPS 2022), where structured covariates are transformed into unstructured proxies. To assess how much the performance gain reflects multimodal reasoning, we conducted a full-tabular oracle that directly observes all 25 original IHDP covariates before their transformation into text and images. Its ϵPEHE\sqrt{\epsilon_{PEHE}} is 1.254, compared with 3.824 for the nine-variable tabular baseline and 3.121 for MultiCFR. Thus, the deterministic multimodal views provide recoverable treatment-effect-relevant information: MultiCFR reduces ϵPEHE\sqrt{\epsilon_{PEHE}} by 18.4% over the restricted tabular baseline and closes 27.4% of the gap to the full-information oracle. The remaining gap shows that recovering information from transformed multimodal views is still much harder than directly observing the original covariates.

Besides, our original MIMIC-TIT experiments use real clinical tabular records, radiology reports, and chest radiographs rather than deterministic transformations. MultiCFR still reduces ϵPEHE\sqrt{\epsilon_{PEHE}} by 35.2% over tabular-only estimation, providing complementary evidence that its effectiveness extends to genuinely heterogeneous multimodal data.

Tab. 5 Capacity-Matched View-Recovery Comparison

ϵPEHE\sqrt{\epsilon_{PEHE}}ϵATE\epsilon_{\mathrm{ATE}}Factual RMSE
Tabular-9 CFR3.824±0.6160.505±0.0833.965±0.570
Tabular-9 + Text CFR3.591±0.5910.534±0.1153.760±0.547
Tabular-9 + Image CFR3.521±0.5760.492±0.0743.632±0.522
Standard tri-modal fusion CFR3.057±0.5170.411±0.0593.186±0.466
MultiCFR3.121±0.5450.430±0.0773.253±0.483
Full-25 tabular CFR (oracle)1.254±0.1730.254±0.0371.681±0.149
AuthorsOpenReview ↗

Response [2 / 2]

Q4&W1: Evaluation on Real Observational Data

R4: Thank you for raising this important point. As you noted, evaluating treatment effects in real-world settings remains an open challenge, since individual-level counterfactual outcomes are unobservable. This limitation is not specific to MultiCFR and is common in treatment-effect estimation. Consequently, synthetic and semi-synthetic datasets with known potential outcomes are widely used for direct ITE evaluation in prior work, such as [1,2,3].

For real observational data, researchers typically rely on indirect metrics, such as factual outcome prediction, ATE/ATT, and policy-risk evaluation [4,5]. Although these metrics cannot directly measure ITE error, they provide complementary evidence on reliable outcome prediction, population-level treatment effects, and treatment decision quality.

[1] Shalit, Uri, Fredrik D. Johansson, and David Sontag. "Estimating individual treatment effect: generalization bounds and algorithms." International conference on machine learning. PMLR, 2017.

[2] Yan, Yuguang, et al. "Matching without Group Barrier for Heterogeneous Treatment Effect Estimation." International Conference on Learning Representations. Vol. 2026. 2026.

[3] Kazemi, Amirreza, and Martin Ester. "Adversarially balanced representation for continuous treatment effect estimation." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 12. 2024.

[4] Guo, Jiayi, et al. "A Relative Error-Based Evaluation Framework of Heterogeneous Treatment Effect Estimators." International Conference on Learning Representations. Vol. 2026. 2026.

[5] Hess, Konstantin, et al. "Overlap-weighted orthogonal meta-learner for treatment effect estimation over time." International Conference on Learning Representations. Vol. 2026. 2026.

W4: Clarification of the Dual-Branch Information Allocation

R5: We clarify that MultiCFR employs soft allocation rather than strict causal disentanglement; thus, ZBZ^B and ZRZ^R are not expected to contain mutually exclusive confounding and predictive information. This design avoids discarding overlapping treatment- and outcome-relevant signals while selectively balancing the information allocated to ZBZ^B.

Following your suggestion, we conducted additional diagnostics. Treatment-prediction AUCs are nearly identical for ZBZ^B and ZRZ^R, confirming that treatment information is not forcibly separated. The joint representation also achieves the highest true-ITE R2R^2 (0.843). Despite this overlap, MultiCFR reduces the weighted residual discrepancy from 0.177 under global balancing to 0.162 (8.5% reduction), while reducing ATE error by 28.1%, supporting the effectiveness of selective balancing.

Tab.7 Dual-Branch Diagnostics

DiagnosticMetricZBZ^BZRZ^RJoint / Result
Treatment PredictionAUC0.6910.690—
Information RetentionPrognostic R2R^20.8750.8720.880
Information RetentionTrue ITE R2R^20.8360.8340.843
Residual DiscrepancyNo Balance——0.172±0.0150.172\pm0.015
Residual DiscrepancyGlobal Balance——0.177±0.0140.177\pm0.014
Residual DiscrepancyMultiCFR——0.162±0.0110.162\pm0.011

Dear Reviewer 5K5i,

We greatly appreciate your time and effort in reviewing our work. We hope that our responses have clarified our work and addressed your concerns. If so, we would sincerely appreciate your consideration of updating your rating accordingly. If you have any further concerns, we would be happy to provide further clarification.

We highly value every exchange during the review process and sincerely hope to receive further feedback from you at your earliest convenience.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.