acceptodds
Under review as a conference paper at ICLR 2027

JETO: Joint Estimation of Treatment Assignment and Outcomes for Heterogeneous Treatment Effect Analysis

Abstract

Heterogeneous treatment effect estimation seeks to characterize how the effect of a treatment varies across individuals with different covariates, despite only one potential outcome being observed for each unit. Classical methods primarily target the conditional average treatment effect, providing a point estimate of treatment benefit. Generative modeling of \(p(Y\mid X,W)\) extends this view to treatment-specific outcome distributions, but leaves the treatment assignment mechanism outside the generative target and does not explicitly capture structure shared between propensity and outcomes. We propose JETO, which learns the joint conditional distribution \(p(Y,W\mid X)\) with conditional Flow Matching, providing a treatment effect estimate, treatment-specific means, propensity scores, and conditional outcome densities from a single model. Across six synthetic and semi-synthetic benchmarks, JETO achieves the lowest mean relative PEHE on four datasets. Controlled experiments show that joint modeling exploits covariate-dependent structure shared by treatment assignment and outcomes, while density experiments demonstrate accurate multimodal outcome recovery and propensity estimation.

Primary area:
causal reasoning
Keywords:
heterogeneous treatment effect, causal inference, conditional Flow Matching, joint modeling, propensity estimation, outcome distributions

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

This paper proposes JETO, a conditional flow-matching model that learns the joint observational distribution p(Y,W∣X)p(Y,W\mid X). A single model provides treatment-specific outcome distributions, conditional average treatment effects, and propensity scores. Across six synthetic and semi-synthetic benchmarks, JETO achieves the lowest mean relative PEHE on four datasets. An architecture-matched comparison shows benefits when treatment propensity and outcome modality share covariate-dependent structure. Additional experiments examine conditional-density recovery, propensity estimation, and numerical inference.

Critical strengths
  1. Competitive empirical performance. JETO achieves the lowest mean relative PEHE on four of six benchmarks, including a substantial improvement on IHDP. These results shows practical promise for treatment-effect estimation.

  2. A useful controlled study. The architecture-matched comparison between joint and conditional modelling provides evidence that the assignment signal can improve estimation when propensity and outcome modality share structure. These findings support a potentially useful empirical contribution.

Critical weaknesses
  1. Insufficient motivation for joint generation. Under the standard identification assumptions, the potential-outcome distribution p(Y(w)∣X)p(Y(w)\mid X) is identified by p(Y∣W=w,X)p(Y\mid W=w,X). Joint generation models observational treatment assignment and outcomes; it does not itself specify an intervention mechanism or provide additional causal identification. Its motivation therefore rests primarily on estimation benefits, which need to be more clearly established relative to separately modeling p(Y∣W,X)p(Y\mid W,X) and p(W∣X)p(W\mid X).

  2. Limited methodological novelty. The method mainly applies standard conditional flow matching to an augmented treatment-outcome target, with treatment encoding or continuous relaxation. I find this extension incremental, and the paper does not sufficiently establish its additional methodological value beyond existing conditional generative approaches.

Questions for authors
  1. Can the authors provide a capacity- and compute-matched comparison against a conditional flow-matching outcome model with a shared covariate representation and an auxiliary propensity-score head? This targeted comparison would materially affect my assessment of the contribution of joint generation.

  2. Does the advantage over architecture-matched conditional JETO persist on IHDP, where the main benchmark reports the largest improvement? A paired comparison would help establish whether that improvement is attributable to the joint formulation.

1 comment

We thank the reviewer for the detailed assessment and for identifying a concrete comparison that could help clarify the contribution of JETO. We would like to emphasize a distinction central to our work: while the joint distribution admits an exact factorization into propensity and conditional outcome distributions, this equivalence does not imply equivalent finite-sample estimation performance for models trained under different objectives. Our contribution lies in investigating the estimation benefits of joint generation, rather than proposing additional causal identification assumptions. To address the reviewer's concerns directly, we provide a new capacity-matched comparison with auxiliary propensity supervision and paired results on IHDP, showing that the benefits of joint generation cannot be explained solely by the additional assignment supervision in the evaluated settings.

Q1

Can the authors provide a capacity- and compute-matched comparison against a conditional flow-matching outcome model with a shared covariate representation and an auxiliary propensity-score head? This targeted comparison would materially affect my assessment of the contribution of joint generation.

Yes. Revised Table 2 adds a conditional outcome flow with a shared covariate representation and an auxiliary propensity head. We matched model capacity by parameter count and used the same total Monte Carlo budget. The table reports paired comparisons on the same realizations. Values below are mean rPEHE, with lower values better. The last column gives the unadjusted paired 95% tt interval for the propensity-head control minus joint JETO.

DatasetJoint JETOConditional JETOTwo-Head (Conditional + Propensity Head)Paired 95% CI (Conditional - Joint)Paired 95% CI (Two-Head - Joint)
IHDP0.15500.16670.1728[-0.0007, 0.0242][0.0060, 0.0295]
GM-I0.50860.49060.5435[-0.0393, 0.0034][0.0046, 0.0653]
GM-P0.58340.62730.6267[0.0180, 0.0699][0.0066, 0.0802]

Joint JETO outperforms the requested propensity-head control on all three datasets, with all three paired intervals above zero. GM-P provides the most direct test of the proposed mechanism. Adding a propensity head to the conditional flow changes mean rPEHE from 0.6273 to 0.6267, whereas joint generation reduces it to 0.5834. On GM-I the outcome-only conditional model has the lowest mean. These results locate the contribution at the level of estimation under the same causal identification assumptions.

Q2

Does the advantage over architecture-matched conditional JETO persist on IHDP, where the main benchmark reports the largest improvement? A paired comparison would help establish whether that improvement is attributable to the joint formulation.

Yes in mean rPEHE, and the advantage also holds against the assignment-supervised control requested by the reviewer. Across all 99 paired IHDP evaluation realizations, joint JETO achieves the lowest mean of the three variants. Its rPEHE is approximately 7% lower than the outcome-only conditional flow’s (0.1550 vs. 0.1667) and 10% lower than the propensity-head control’s (0.1550 vs. 0.1728). For conditional minus joint, the paired 95% interval is [−0.0007, 0.0242], leaving a near-tie compatible with the data. For the propensity-head control minus joint, the interval is entirely positive [0.0060, 0.0295]. Thus the favorable IHDP result survives a within-architecture comparison and is clearest against the very control proposed to explain it. The full comparison with external baselines and this paired ablation answer complementary questions about overall performance and the contribution of the joint formulation.

Reviewer_ZLnHOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes JETO, which investigates whether jointly modeling treatment assignment and outcomes can improve heterogeneous treatment-effect estimation and recover richer distributional information. Whereas conditional outcome generators take treatment as an input, JETO uses conditional Flow Matching to learn the joint distribution of treatment and outcome given observed covariates. Its proposed benefit is the additional assignment-learning signal when propensity and outcome distributions share covariate-dependent structure. Under standard causal identification assumptions, treatment-specific means, treatment effects, and propensity can be estimated from joint samples. A separate Gaussian-relaxed treatment representation enables conditional-density evaluation. Across six synthetic and semi-synthetic benchmarks, JETO achieves the lowest mean relative PEHE on four datasets. An architecture-matched comparison finds a benefit from joint modeling in a constructed setting where outcome-mixture weights depend on propensity. Further experiments are also provided, examining multimodal density recovery, propensity estimation and other aspects, providing evidence about both the accuracy and implementation of the joint generative approach.

Critical strengths
  • The paper considers a meaningful and interesting problem. Estimating how treatment effects vary across individuals is an important causal-inference task, and learning outcome distributions provides information beyond average effects. The paper asks whether information about treatment assignment can improve this estimation when assignment and outcomes share covariate-dependent structure. In my humble opinion, this is a worthwhile question because it connects propensity learning with the quality of treatment-effect and distributional estimates within a common model.

  • The experiments are overall supportive. The evaluation covers six synthetic and semi-synthetic benchmarks, with separate tuning and evaluation realizations and equal hyperparameter-search budgets. JETO achieves the lowest mean rPEHE on four datasets. The controlled comparison in Table 2 also supports a benefit from joint modeling on GM-P, with a paired rPEHE difference of 0.0440 and a 95% interval of [0.0180, 0.0699] favoring JETO. Density-recovery experiments and checks of sampling budgets and numerical solvers provide further support for the approach within the tested settings.

Critical weaknesses
  • The technical novelty requires further justification. Existing works such as PO-Flow already uses conditional Flow Matching for potential-outcome estimation and density evaluation, Dragonnet combines outcome and propensity supervision, and CEVAE jointly models treatment and outcome under a different latent-variable formulation. It seems that JETO's main change is to include treatment in the generated target while retaining a standard Flow Matching objective. This is a useful extension, but the comparison in Section 6.2 does not isolate its benefit from that of auxiliary propensity supervision because the conditional control lacks that objective. The evidence therefore does not clearly establish an advantage specific to generating treatment jointly.

  • The paper provides no theoretical guarantee for JETO's estimation accuracy or generalization. Section 3 states standard causal identification assumptions, but these do not establish the accuracy of the learned flow or the resulting CATE estimator. The paper gives no consistency result, error bound, or sufficient conditions under which joint modeling improves on conditional modeling. In particular, it does not explain how errors in the learned joint distribution affect conditional-mean estimates when one treatment arm has low probability. Similar gradient magnitudes and small gradient correlations do not establish such a guarantee. This leaves the conditions under which the proposed learning signal should improve treatment-effect estimation insufficiently understood.

Questions for authors
  1. In Section 6.2, which comparison isolates the benefit of joint generation from auxiliary propensity supervision? Further clarification of the conditional baseline's supervision is suggested because this determines whether the experiment supports the claimed methodological advance.

  2. For Equations (8) and (9), how does error in the learned joint distribution affect CATE error when one treatment arm is rare? Clarifying the required assumptions would help assess the estimator's reliability beyond standard causal identification.

  3. JETO is outperformed by the X-learner on ACIC and the R-learner on NW-B in Table 1. What do these results imply about its strengths and limitations, and when should JETO be preferred in practice?

2 comments

We thank the reviewer for the thoughtful comments and constructive suggestions. Below, we clarify the methodological contribution of joint generation, provide additional experimental evidence distinguishing it from auxiliary propensity supervision, discuss the effects of distributional and sampling errors on CATE estimation, and outline the practical strengths and limitations of JETO.

Q1

In Section 6.2, which comparison isolates the benefit of joint generation from auxiliary propensity supervision? Further clarification of the conditional baseline's supervision is suggested because this determines whether the experiment supports the claimed methodological advance.

The original conditional control received treatment as an input and had only an outcome flow-matching loss. Revised Table 2 adds a conditional outcome flow with a shared covariate representation and an auxiliary propensity head. Comparing joint JETO with this third model directly addresses the reviewer’s question: both learn from assignment labels and have comparable parameter counts. On GM-P the auxiliary head yields mean rPEHE 0.6267, close to the outcome-only model’s 0.6273, while joint JETO achieves 0.5834. The unadjusted paired 95% interval for two-head minus joint is [0.0066, 0.0802]. The gain therefore persists when the control also receives propensity supervision. We use standard flow matching. The methodological contribution is the joint generative formulation and the controlled test of what that choice adds beyond a shared representation with assignment supervision.

DatasetJoint JETOConditional JETOTwo-Head (Conditional + Propensity Head)Paired 95% CI: Conditional and Joint DifferencePaired 95% CI: Two-Head and Joint difference
IHDP0.15500.16670.1728[-0.0007, 0.0242][0.0060, 0.0295]
GM-I0.50860.49060.5435[-0.0393, 0.0034][0.0046, 0.0653]
GM-P0.58340.62730.6267[0.0180, 0.0699][0.0066, 0.0802]

Q2

For Equations (8) and (9), how does error in the learned joint distribution affect CATE error when one treatment arm is rare? Clarifying the required assumptions would help assess the estimator's reliability beyond standard causal identification.

Fix xx and let pw=P(W=w∣x)p_w=P(W=w\mid x) and mw=E[Y 1{W=w}∣x]m_w=E[Y\,\mathbf{1}_{\{W=w\}}\mid x]. Write qwq_w and m~w\tilde m_w for the same quantities under the learned distribution. Assuming finite first moments and positive arm probabilities, μw=mw/pw\mu_w=m_w/p_w and the exact error is

μ~w−μw=(m~w−mw)−μw(qw−pw)qw.\tilde\mu_w-\mu_w=\frac{(\tilde m_w-m_w)-\mu_w(q_w-p_w)}{q_w}.

A small qwq_w can amplify error in either the estimated arm mass or its outcome-weighted first moment. The CATE error is the difference of the two arm-mean errors. Convergence of both masses and first moments gives pointwise convergence of the model-implied CATE wherever both true arm probabilities are positive. A convergence rate for the fitted flow would require additional assumptions.

Finite sampling adds a separate source of error. With KK independent draws, the expected arm count is KqwKq_w and the empty-arm probability is (1−qw)K(1-q_w)^K. If either arm is empty, the implementation returns the estimated ATE for that observation. The lower-right panel of the appendix Monte Carlo budget figure measures this count in poor-overlap regions (e(x)<0.1e(x)<0.1 or e(x)>0.9e(x)>0.9). At K=2048K=2048, the generated minority-arm count averages roughly 220 on GM-I and 280–290 on GM-P and NW-D. These are averages over the selected observations. The GM-P density experiment provides a complementary check: for treated outcomes near e(x)=0.1e(x)=0.1, conditional-density KL is 0.810 for JETO versus 1.219 for PO-Flow.

Q3

JETO is outperformed by the X-learner on ACIC and the R-learner on NW-B in Table 1. What do these results imply about its strengths and limitations, and when should JETO be preferred in practice?

JETO achieves the lowest mean rPEHE on four of six benchmarks. On NW-B, randomization gives e(x)=0.5e(x)=0.5, so treatment assignment supplies no covariate-dependent signal. The R-learner achieves 0.331 versus 0.420 for JETO. On ACIC the X-learner achieves 0.142 versus 0.218, and the benchmark comparison does not identify the cause of that gap. We see the strongest case for JETO when the application requires treatment-specific outcome distributions alongside CATE, especially when assignment and outcome heterogeneity plausibly depend on common predictors. For a CATE-only application with tight computational budgets, the boosting-based learners remain strong candidates.

Thank you for the detailed response. The added propensity-supervised comparison and error analysis help address my original concerns. I retain some reservations as the guarantee on CATE estimation accuracy also remains absent, particularly when one treatment arm is rare. The error decomposition usefully explains this sensitivity, but the reliability of the learned estimator in such settings remains uncertain.

Reviewer_iM9nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes JETO, a conditional generative model that learns the joint distribution p(Y,W | X) using Flow Matching, allowing a single model to estimate CATE, treatment-specific outcome distributions, and propensity scores. The method is evaluated on six synthetic and semi-synthetic benchmarks and is compared with classical meta-learners, CEVAE, DiffPO, and PO-Flow.

Critical strengths
  1. The idea is straightforward and clear. Jointly modeling treatment assignment and outcomes provides a unified generative formulation for treatment effects, outcome densities, and propensity scores.
  2. The matched GM-I/GM-P experiments directly test when joint modeling helps and show an advantage when treatment assignment and outcome modality share covariate-dependent structure.
  3. The paper includes multiple HTE baselines, density-recovery experiments, propensity estimation, architecture-matched ablations, gradient analyses, and inference sensitivity studies.
Critical weaknesses
  1. The empirical validation is entirely synthetic or semi-synthetic. The main advantage of JETO is therefore demonstrated only in settings where the true CATE and data-generating mechanism are known. It remains unclear whether joint modeling provides a meaningful advantage in real observational applications.
  2. The benefit of joint modeling appears setting-dependent. JETO is not best on ACIC or NW-B, and the architecture-matched comparison shows a clear advantage only in GM-P.
  3. The practical value of learning propensity jointly is not fully established. The authors acknowledge that propensity modeling is not required for CATE identification under their assumptions. The added benefit therefore seems primarily representational, and it is unclear how often the required shared assignment–outcome structure occurs in real datasets.
  4. Inference is relatively expensive. JETO relies on ODE integration and large Monte Carlo sampling budgets, with K=2048 samples per covariate vector in the main experiments, which may limit scalability.
Questions for authors
  1. Can the authors provide a real-world observational application, even if true CATE is unavailable, to demonstrate that the joint formulation yields useful downstream information beyond synthetic benchmarks?
  2. How common do the authors expect the GM-P-type setting to be in practice, where treatment assignment and outcome modality share sufficiently strong covariate-dependent structure for joint modeling to help?
  3. Since propensity estimation is not necessary for CATE identification here, can the authors better characterize when the additional treatment-assignment objective is expected to improve estimation rather than simply add modeling complexity?
  4. How does the computational cost of JETO compare with PO-Flow, DiffPO, and standard HTE estimators at inference time?
1 comment

We thank the reviewer for the constructive feedback. Below, we clarify the motivation and practical scope of JETO, discuss when joint modeling can be beneficial, and present additional experimental evidence on its estimation performance and computational cost.

Q1

Can the authors provide a real-world observational application, even if true CATE is unavailable, to demonstrate that the joint formulation yields useful downstream information beyond synthetic benchmarks?

Evaluating CATE accuracy against known effects on synthetic and semi-synthetic benchmarks is standard practice in treatment-effect research. IHDP (Hill, 2011) and ACIC 2016 (Dorie et al., 2019) use real covariates and simulated outcomes, allowing every estimator to be evaluated against the same reference effect. A fully observational application could illustrate the use of JETO’s treatment-specific predictive distributions. Its factual outcomes alone would leave open the question central to this paper: which formulation estimates CATE more accurately? We therefore focus our additional evaluation on the paired comparison requested by the reviewers, including all 89 IHDP evaluation realizations.

Q2

How common do the authors expect the GM-P-type setting to be in practice, where treatment assignment and outcome modality share sufficiently strong covariate-dependent structure for joint modeling to help?

Consider a physician who prescribes a therapy more often to patients with a recorded pretreatment biomarker associated with a favorable response. The physician need only know that response is more likely in this subgroup. Both assignments and treated outcomes then provide information about the same patient characteristic. GM-P captures this mechanism by making the treated-outcome mixture weight depend on propensity. Its prevalence across populations is an empirical question. Our controlled experiment establishes that it can produce a measurable estimation benefit when present.

Q3

Since propensity estimation is not necessary for CATE identification here, can the authors better characterize when the additional treatment-assignment objective is expected to improve estimation rather than simply add modeling complexity?

Under the considered identification assumptions the conditional outcome law suffices to identify CATE. The proposed benefit arises in finite samples when learning assignment helps recover variation in the outcome distribution. Revised Table 2 tests this explanation. On GM-P, joint JETO achieves mean rPEHE 0.5834, versus 0.6273 for conditional JETO and 0.6267 after adding a propensity head. Both paired intervals favor joint generation. On GM-I, where mixture weights are independent of propensity, the outcome-only conditional model has the lowest mean. Thus predictable assignment alone is insufficient to expect a gain.

DatasetJoint JETOConditional JETOTwo-Head (Conditional + Propensity Head)Paired 95% CI: Conditional and Joint DifferencePaired 95% CI: Two-Head and Joint difference
IHDP0.15500.16670.1728[-0.0007, 0.0242][0.0060, 0.0295]
GM-I0.50860.49060.5435[-0.0393, 0.0034][0.0046, 0.0653]
GM-P0.58340.62730.6267[0.0180, 0.0699][0.0066, 0.0802]

Q4

How does the computational cost of JETO compare with PO-Flow, DiffPO, and standard HTE estimators at inference time?

We measured wall-clock inference time per test split on all six benchmarks (see the anonymized comparison). With the same total budget of K=2048K=2048 trajectories per covariate vector, JETO and PO-Flow have closely comparable mean runtimes on every dataset. The original DiffPO implementation was the slowest on all six under the evaluated settings. CEVAE was faster than either flow, and boosting-based HTE estimators were faster by orders of magnitude. For CATE alone, this cost matters. When outcome distributions are also needed, JETO provides them at an observed inference cost close to PO-Flow’s.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.