acceptodds
Under review as a conference paper at ICLR 2027

Neighborhood-Guided Flow Matching for Potential Outcome Distributions

Abstract

Recent work in causal inference has increasingly focused on modeling conditional distributions of potential outcomes, which capture outcome variability beyond conditional means. Balanced representation learning aims to mitigate selection bias in observational data, but representation alignment alone does not ensure that fine outcome variations are preserved. To address this challenge, we propose a unified framework that couples representation balancing with neighborhood-guided flow matching. Our method encourages alignment between treatment groups in latent space and uses observed outcomes from covariate neighbors in the target treatment arm as flow-matching endpoints, while conditioning the velocity field on the individual's learned representation and the target treatment. By weighting neighborhood supervision according to local treatment support, the model borrows outcome information where suitable neighboring observations are available. Under standard identification and smoothness assumptions, we show that neighborhood supervision targets a local mixture of conditional potential outcome distributions and bound its deviation from the true distribution. Experiments demonstrate the effectiveness of our method in conditional distribution recovery on synthetic data and treatment-effect estimation on standard causal benchmarks.

Primary area:
causal reasoning
Keywords:
causal inference, flow matching, potential outcome distributions, representation learning, selection bias, local supervision, observational Data

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

This paper addresses an important problem in causal inference: learning the full conditional distributions of potential outcomes, rather than only their conditional means, when observational treatment assignment creates uneven local support. Its proposed method, Neighborhood-Guided Flow Matching (NGFM), combines representation balancing, local cross-treatment donor selection, and conditional flow matching. The key idea is to use outcomes from nearby observations in the opposite treatment arm as weighted generative endpoints, while reducing the influence of donors when local cross-arm support is poor. The paper targets the conditional potential-outcome distributions. Under consistency, conditional exchangeability, and positivity, these distributions are identified from the corresponding factual treatment groups. In finite samples, however, the target treatment arm may have few observations near a query point. NGFM has three principal components, Representation learning and balancing, A shared encoder learns covariate representations, while a debiased Sinkhorn divergence encourages alignment between treatment groups. A frozen copy of the encoder, obtained after factual warm-up, defines a stable neighborhood graph. Neighborhood-guided local supervision, For each query, the method selects opposite-arm donors. Gaussian distance weights q_ij distribute influence among donor outcomes, while a reliability weight r_i reduces the overall auxiliary loss when cross-arm donors are distant relative to the local pooled data density. Distributional flow matching, Each donor outcome is used as an endpoint for a flow-matching path conditioned on the query's representation and the opposite treatment. The overall objective combines factual flow matching, representation balancing, and neighborhood-guided supervision. At inference time, the fitted flow generates samples without retrieving neighbors. The paper evaluates distributional recovery on synthetic settings and HC-MNIST, and effect estimation on IHDP and ACIC 2018. Its theoretical analysis characterizes the donor-induced endpoint mixture and bounds the discrepancy between that mixture and the target distribution.

Critical strengths
  1. It addresses distributional information that mean-effect methods miss. The full potential-outcome distribution can reveal differences in variance, tail risk, and multimodality even when mean outcomes are similar. This is particularly relevant when treatment decisions depend on the probability of an adverse outcome rather than only the expected response. NGFM's choice to retain multiple donor outcomes as separate flow endpoints is well motivated. The paper's synthetic examples illustrate a real failure mode: a generator may approximate the conditional mean yet smooth away separated modes.
  2. The reliability and donor weights serve distinct purposes. The distinction between r_i and q_ij is one of the strongest parts of the design. r_i determines the total strength of borrowed supervision for a query. q_ij determines the allocation of that supervision among the selected donors. This decomposition is interpretable and could be useful beyond flow matching. It also makes it easier to investigate whether an improvement comes from better donor selection, better weighting, or simply more auxiliary training data.
  3. The paper explicitly analyzes the target induced by the training objective. The endpoint-mixture characterization is a meaningful strength. Rather than claiming that neighborhood supervision automatically recovers the target distribution, the paper identifies a mixture between the true conditional law and the empirical donor law. Its subsequent bounds relate donor approximation error to neighborhood distance, effective donor count, and flow approximation error. This is a useful step toward understanding the statistical consequences of local supervision. It is more informative than an empirical proposal supported only by benchmark improvements.
Critical weaknesses
  1. The core causal question: are nearby factual outcomes valid supervision for the missing potential outcome? The central assumption behind NGFM is that observations close in the learned representation have similar conditional potential-outcome distributions. The paper formalizes this through a smoothness condition. This is a reasonable assumption for local distributional estimation, but it is not guaranteed by representation balancing or predictive training. Two individuals may be close in the learned space but differ in a clinically or causally important effect modifier. Conversely, two individuals far apart in the representation may have similar potential-outcome distributions. Global alignment of treatment groups does not establish that the representation is sufficient for preserving the outcome distributions required for causal estimation. This issue is especially important under selection bias. If the treatment assignment mechanism strongly depends on covariates, the nearest opposite-arm donor may still be a poor proxy for the query's potential outcome. The reliability score detects one form of local support scarcity, but it cannot detect all forms of representation misspecification or unmeasured effect modification.
  2. The theoretical guarantee has a significant gap between its assumptions and the implemented procedure. The paper's donor approximation bound assumes that, conditional on the graph-defining information, donor outcomes are independent and have the relevant conditional laws. In the appendix, the authors explicitly state that the concentration bound additionally assumes a metric fitted independently of donor outcomes, whereas the implemented factual warm-up trains the metric on those same outcomes. The authors acknowledge that the warm-up procedure does not establish the required independence. This is an important limitation, not a minor technicality. The learned metric determines which outcomes are selected as donors, so dependence between metric construction and donor outcomes can affect the validity of the concentration argument. The empirical endpoint-law identity may still hold for the implemented objective, but the stronger population approximation guarantee does not automatically follow for that implementation.
  3. Factual validation CRPS does not directly validate counterfactual distribution recovery. The paper selects checkpoints using factual validation CRPS. CRPS is a proper scoring rule for distributional forecasts, which is appropriate for evaluating a predictive distribution. However, in observational data, validation outcomes are observed only under the treatment actually received. A model may achieve good factual CRPS while estimating the unobserved potential-outcome distribution poorly in regions with weak overlap. Factual validation therefore does not directly select the model with the best counterfactual distributional recovery. This is a particularly consequential issue because the main contribution is supposed to improve the distribution of the missing potential outcome.
Questions for authors
  1. The donor approximation theorem assumes a metric fitted independently of donor outcomes, but the implementation fits the warm-up encoder on those outcomes. Can the authors prove a suitable bound for the implemented procedure, or provide cross-fitting experiments that satisfy the independence assumption?
  2. What evidence shows that nearby points in the learned representation have similar conditional potential-outcome distributions, rather than merely similar factual outcome predictions? Can this be tested directly in simulation?
  3. Checkpoints are selected using factual validation CRPS. How can the authors establish that this criterion selects models with good counterfactual distributional recovery in observational data, where counterfactual outcomes are unobserved?
Reviewer_HGTaOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Neighborhood-Guided Flow Matching (NGFM) for estimating conditional marginal distributions of potential outcomes from observational data. The method combines factual flow-matching supervision with representation balancing and an auxiliary loss constructed from observed outcomes of nearby units in the opposite treatment group. A representation learned during factual warm-up is frozen for neighborhood construction, while the prediction model continues to train. Distance-based donor weights and a query-specific reliability factor control the contribution of the borrowed outcomes. The theoretical analysis characterizes the resulting factual–donor endpoint mixture and relates its distributional error to neighborhood approximation, donor sampling, and flow approximation errors, with additional independence assumptions required for the donor concentration bound. Experiments cover synthetic distribution-recovery tasks, HC-MNIST, and treatment-effect estimation on IHDP and ACIC 2018, supplemented by component ablations and sensitivity analyses. The central contribution is the use of weighted local outcome distributions, rather than only local mean targets, to provide additional supervision under uneven treatment support.

Critical strengths
  1. The paper addresses a meaningful limitation of treatment-effect estimation based primarily on conditional means. Its motivation is particularly relevant when outcome distributions differ in shape despite having similar means, and when treatment support is uneven across the covariate space. This gives the work a clear substantive purpose and makes distributional recovery an appropriate focus of the evaluation.

  2. The use of neighboring observed outcomes as additional flow-matching endpoints is a plausible and interpretable idea. Preserving the collection of donor outcomes allows local supervision to convey information about distributional shape, while the reliability mechanism acknowledges that borrowing is less trustworthy in poorly supported regions. The separation between neighborhood construction during training and prediction without neighbor retrieval also gives the method a practical appeal.

Critical weaknesses
  1. The connection between the theoretical motivation and the learned neighborhoods needs stronger empirical support. The paper states assumptions that justify local borrowing, but it remains unclear how well the learned representation preserves the similarities in outcome distributions that this reasoning requires. Although component ablations are included, aggregate performance differences offer limited insight into whether the selected donors are appropriate and whether reliability weighting effectively reduces harmful borrowing. This limits confidence in the central mechanism and its applicability beyond the reported examples.

  2. The experimental presentation does not sufficiently explain the scope of the reported benefits. Performance varies across settings and evaluation criteria, but the implications of these differences need clearer discussion. In Figure 2, both panels are titled “Setting I,” without identifying the treatment arms or adequately distinguishing the examples. Both convey essentially the same qualitative message, while their selection and representativeness remain unclear, limiting the information gained relative to the space occupied. The method schematic is also densely annotated, with some labels difficult to inspect. Clarification of how it was prepared, together with an original editable source or a clear vector version, would help readers examine its content and correspondence to the method.

Questions for authors
  1. What evidence shows that proximity in the learned representation corresponds to similarity in conditional outcome distributions? What do the experiments establish about the ability of reliability weighting to reduce harmful borrowing when the learned neighborhoods are imperfect?

  2. Which experimental findings establish the conditions under which neighborhood guidance helps, and which reveal its limitations? A clearer interpretation of the differences across settings and metrics would help assess the scope of the contribution.

  3. How were the two query–arm pairs in Figure 2 selected, and which treatment arm does each represent? Please explain their representativeness and what distinct information each panel contributes.

Reviewer_Nd6vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Neighborhood-Guided Flow Matching (NGFM) for estimating conditional potential outcome distributions. It combines representation balancing with flow matching and uses weighted outcomes from neighboring samples in the opposite treatment group as additional supervision. Theoretical analysis and experiments on synthetic and semi-synthetic datasets demonstrate improvements in distribution recovery and treatment effect estimation in several settings.

Critical strengths
  1. The paper addresses an important problem in causal inference, focusing on recovering full potential outcome distributions rather than only average treatment effects.
  2. The proposed neighborhood-guided supervision is intuitive and helps preserve local outcome variations.
  3. The paper provides theoretical analysis and extensive experiments, including comparisons with existing methods, ablation studies, and parameter sensitivity analysis.
Critical weaknesses
  1. The theoretical guarantee in Section 4 assumes an outcome-independent neighborhood metric, but the encoder is trained on donor outcomes. A sample-splitting experiment is needed to validate the theory.
  2. The novelty over existing methods is unclear. Comparisons with KNN-based distribution estimation and PO-Flow using simple neighbor supervision would help establish the contribution.
  3. Table 1 shows inconsistent improvements. For example, NGFM achieves 0.267 rPEHE in Setting I, compared with 0.106 for DKLITE. The authors should explain why better distribution recovery does not improve treatment effect estimation.
  4. Table 5 compares variants with different preprocessing settings. All variants should be rerun using identical preprocessing to ensure fair comparisons.
  5. Figure 5 shows that increasing K from 4 to 32 worsens cW1 from 0.372 to 0.584 in Setting III. The authors should investigate adaptive neighbor selection and report runtime and memory costs.
Questions for authors

see weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.