acceptodds
Under review as a conference paper at ICLR 2027

Generalized R-learner for Estimating Potential Outcome and Treatment Effects

Abstract

Estimating potential outcome and treatment effects from observation data is important for decision making. Orthogonal learners represent a growing area of research that develops target estimators insensitive to plug-in nuisance functions. Among existing orthogonal learners, the R-learner offers an approach that is free from numerical instability due to inverse propensities, yet it is derived based on partially linear outcome equations. We develop a *target-error* analysis for the R-learner to isolate an explicit *structural-drift* term that represents the gap between class-optimal plug-in target and the oracle target, revealing that the partially linear structure leads to a structural-drift *not reducible by the target estimator.* We accordingly present a generalized R-learner (GRL) that is able to accommodate this structural drift by utilizing a general nonlinear outcome equation in the target estimator, and provide an orthogonality regularized objective function that serves the dual purpose of reducing nuisance-induced structural drift in the target estimator while providing a nuisance-robust population risk bound. We establish conditions under which GRL achieves smaller target errors and when the target remains resolvable under nuisance estimation errors. On (semi-)synthetic and real datasets, we empirically demonstrated the strength of GRL in comparison to existing meta-learners in both potential outcome and treatment effect estimation.

Primary area:
causal reasoning
Keywords:
Causal Inference, Treatment Effect, Orthogonal-learners, Meta-learners

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tLXPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper analyzes structural drift in the R-learner caused by nuisance estimation errors. It proposes generalized R-learner, which replaces the partially linear second-stage model with a nonlinear MLP and adds an orthogonality regularization loss. The method is evaluated on IHDP, Twins, and Jobs data.

Critical strengths

The paper studies an interesting problem where nuisance estimation errors can cause structural drift that standard R-learner cannot fully accommodate. The proposed method is intuitive and straightforward, and is supported with both theoretical analysis and experiments on common CATE estimation benchmarks.

Critical weaknesses
  1. While the method includes an orthogonality penalty, it no longer retains the exact Neyman orthogonality property of the R-learner. It seems to me that the paper introduces a trade-off where the more flexible second-stage model can better account for structural drift, but it loses exact orthogonality in return.

  2. The meta-learners used for benchmarking and discussed in the related work are mostly methods proposed before 2023, while several improved variants have been published in the past two years. E.g:

[1] Liang, Zhongyuan, Lars van der Laan, and Ahmed Alaa. "Hybrid Meta-learners for Estimating Heterogeneous Treatment Effects."

[2] Nagalapatti, Lokesh, et al. "PairNet: Training with observed pairs to estimate individual treatment effect."

[3] Chauhan, Vinod Kumar, et al. "Dynamic inter-treatment information sharing for individualized treatment effects estimation."

The paper would benefit from benchmarking against more recent variants to provide a more faithful comparison with current SOTA methods.

  1. The empirical performance gains compared to the baselines seem to be small, particularly on Twins and Jobs.
Questions for authors

Please see my comments in the Weaknesses section. It would be helpful if the authors could revise the paper to include comparisons with these more recent variants, as they have been shown to outperform the meta-learners currently used as baselines in this paper.

3 comments

We thank reviewer for the review and constructive feedback.

Neyman orthogonality (G3 of overall response).
We agree that GRL does not inherit Neyman orthogonality automatically. For R-learner, orthogonality follows from its partially linear structure, but for a general nonlinear hh, the pathwise derivatives in Eq. 13 need not vanish. The penalty encourages small nuisance gradients, but it can make them vanish only if the second-stage class contains an oracle-correct predictor with this property. We have added this condition to the paper (Assumption 12). Under it, Corollary 1 (Appendix C.9) shows that the minimizer of our orthogonality regularized objective (Equation 14) under oracle nuisances is oracle-correct with vanishing nuisance gradients, such that the estimating function is Neyman-orthogonal at this minimizer and the bound in Proposition 7 becomes second order in the nuisance errors. Without this condition, the reviewer's description of a trade-off is accurate: GRL may lose orthogonality and incur some bias in exchange for the flexibility to absorb structural drift that the partially linear class cannot. We now state this trade-off in the Limitations (Section 7).

Addition of more recent baselines (G4 of overall response).
We thank the reviewer for these references. We have added them to the discussion of related works (Section 2) and as baselines (Section 6, 6.1, & Tables 1-2) . We added all three suggested baselines: HyperITE [3] with TARNet backbone, H-learner [1] with X-learner's pseudo-outcome, and PairNet [2]. On synthetic data, PairNEt achieved the lowest PEHE for all values of λ\lambda, but had considerabely higher outcome MSE as its pair loss promotes positive correlation between treatment arms which cancel in CATE, but not in potential outcomes. The H-learner achieved lower PEHE than GRL at only higher values of λ\lambda (λ≥0.8\lambda \ge 0.8), but had higher outcome MSE which also fluctuated across different values of λ\lambda. HyperITE outperformed TARNet for λ≥0.2\lambda \ge 0.2, but remaind behind GRL in both PEHE and Outcome MSE. On real benchmarks, our earlier obserations remain unchanged: GRL had lowest PEHE on IHDP, while differences on Twins and Jobs remained small and not statistically significant across top-performing methods (PairNet achieved lowest policy risk on Jobs).

Performance on Twins and Jobs.
We agree that the differences among top-performers on Twins and Jobs are small and mostly not statistically significant, as we noted in Section 6.2 and the Limitations. We also would like to bring the reviewer's attention to the fact that there was not any method, among the extensive suite of baselines considered, that showed statistically significant improvements over GRL or the next best competitors throughout all datasets. In fact, GRL was the only method that remained among top performers on all three real datasets: top performer with statistically significant difference over the next best competitor on IHDP, and no statistically significant difference to the top performer on Twins and Jobs. In comparison, other baselines fluctuated: e.g., DragonNet is strong on Twins but among the weakest on IHDP and Jobs; the DR-learner and the newly added PairNet showed the reverse pattern. This was the consistency, rather than winning every dataset, that we emphasized in Section 6.2, which is consistent with GRL's expected robustness to nuisance estimation errors that are often unpredictable in real-world datasets.

References:

[1] Liang, Zhongyuan, Lars van der Laan, and Ahmed Alaa. "Hybrid Meta-learners for Estimating Heterogeneous Treatment Effects."

[2] Nagalapatti, Lokesh, et al. "PairNet: Training with observed pairs to estimate individual treatment effect."

[3] Chauhan, Vinod Kumar, et al. "Dynamic inter-treatment information sharing for individualized treatment effects estimation."

Thank you to the authors for the clarifications and additional baselines. I think the revised experiments now provide a more faithful comparison with recent state-of-the-art methods.

I also agree with reviewer g8bU that the R-learner is primarily designed for CATE estimation rather than PO estimation. Similarly, other methods, such as the H-learner, may intentionally sacrifice PO estimation accuracy in certain settings to achieve better PEHE. Therefore, I do not think that poor PO estimation performance necessarily indicates a limitation of these methods, particularly when their PEHE is comparable to that of the proposed GRL.

Despite this remaining concern, I appreciate the authors' efforts in revising the paper and I think the current version represents a meaningful improvement over the original submission. I have therefore decided to increase my score to weak accept.

We thank the reviewer for finding our revision a meaningful improvement over the original submission. We would like to bring the reviewer's attention to the two additional revisions added in response to new discussion (outlined in Overall Response O1 and O2), which we hope further strengthen the paper in two aspects related to the reviewer's recent feedback: 1) revisions in O1 now more precisely delineate when and why the structural drift impacts CATE vs. PO estimation differently, and 2) the addition in O2 provides additional empirical evidence for the trade-off between PO vs. CATE estimation, which is consistent with the reviewer's observation that different methods may prioritize one over the other (and thus different methods may be beneficial depending on the use scenarios).

Reviewer_g8bUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The work proposes a generalized R-learner (GRL) for potential outcome and CATE estimation. It argues that there is a partially linear structure in the standard R-laerner which may introduce a "structural drift" under misspecified nuisances, which does not vanish in the second stage estimator. Specifically, the authors i) derive an error decomposition in order to separate estimation error from this structural drift, ii) propose a second-stage learner with an regularized objective and iii) provide a population risk bound for their estimator. Finally, the proposed learner is validated on synthetic, semi-synthetic and real world datasets.

Critical strengths
  • The paper is mostly self-contained and provides the proofs in the appendix. The theory is quite extensive, but I have strong concerns about the implications and the interpretation (see weaknesses below).

  • The empirical results look good (especially on synthetic data and IHDP). The authors use a large set of existing meta-learners as baselines.

Critical weaknesses
  • My main concern is the theoretical motivation of the paper, and especially whether the claimed limitation of the R-learner actually exists (I am very certain it does not). The authors argue that the "partially linear" structure of the R-learner is not sufficient to be robust against nuisance estimation errors. However, for binary treatments, the decomposition E[Y∣X,T]=l0(X)+(T−e0(X))τ0(X)\mathbb{E}[Y\mid X,T]=l_0(X)+(T-e_0(X))\tau_0(X) is an exact identity and does not depend on any estimator error, linearity assumption, or else. To make this point more clear: This decomposition is independent of whether the true outcome functions are nonlinear or linear. To be fair, the authors acknowledge this, but I do not think they make a clear difference between potential-outcome estimation and CATE estimation in what follwos. Morzywolek et al. (2024), which is not referenced, show that the R-learner can be understood as an overlap-weighted DR-learner within a family of Neyman-orthogonal CATE learners. I therefore do not see why the partially linear exact identity itself should be a limitation of the R-learner. The linear decomposition is exact, and several papers show that neyman-orthogonality holds.

  • This is directly related to the first weakness. I do not agree that the proposed structural drift is a real limitation in CATE estimation. In Proposition 3, the leading structural drift term is SPL≈E[(Δl(X)−τ0(X)Δe(X))2]S_{\mathrm{PL}}\approx\mathbb{E}[(\Delta l(X)-\tau_0(X)\Delta e(X))^2]. However, minimizing the population R-loss gives

τR∗(X)−τ0(X)=Δe(X)(Δl(X)−τ0(X)Δe(X))e0(X)(1−e0(X))+Δe(X)2.\tau_R^*(X)-\tau_0(X)= \frac{\Delta e(X)(\Delta l(X)-\tau_0(X)\Delta e(X))} {e_0(X)(1-e_0(X))+\Delta e(X)^2}.

Thus, the CATE error is second-order in nuisance estimation errors under overlap. In particular, if e^=e0\hat e=e_0 but l^≠l0\hat l\neq l_0, the population R-learner recovers the exact CATE. In contrast, the proposed structural drift remains SPL=E[Δl(X)2]>0S_{\mathrm{PL}}=\mathbb{E}[\Delta l(X)^2]>0. Therefore, while the structural drift can (perhaps) imply an error in PO estimation, this does not necessarily lead to an error in CATE estimation. Therefore, the structural drift does not expose a limitation of the original R learner.

  • My other main concern is the proposed orthogonality-regularized objective in Equation 14. The regularization does not induce Neyman orthogonality but is just a heuristic sensitivity penality to nuisance errors. What is even more important is that it can introduce bias even when the nuisance functions are perfectly estimated.

    COUNTEREXAMPLE: Let XX be constant, T∼Bernoulli(1/2)T\sim\mathrm{Bernoulli}(1/2), and Y=l0+τ0(T−e0)+ξyY=l_0+\tau_0(T-e_0)+\xi_y, where e0=1/2e_0=1/2, τ0≠0\tau_0\neq 0, and E[ξy∣T]=0\mathbb{E}[\xi_y\mid T]=0. Assume oracle nuisance functions η^=η0\hat\eta=\eta_0 and consider the second-stage model h(X,l,T−e)=l+τ(T−e)h(X,l,T-e)=l+\tau(T-e), which is a special case of their model class. Then, Eq. 14 reduces to

    LGRL−OP(τ)=14(τ0−τ)2+14τ2+C,\mathcal{L}_{\mathrm{GRL-OP}}(\tau) =\frac14(\tau_0-\tau)^2+\frac14\tau^2+C,

    where CC does not depend on τ\tau. Therefore, the minimizer is τGRL∗=τ0/2≠τ0\tau^*_{\mathrm{GRL}}=\tau_0/2\neq\tau_0.

    Thus, even with oracle nuisance functions and a correct second-stage model, the regularization can introduce nonvanishing bias. This is highly problematic, because the main claim of the paper is that it generalizes the R-learner. However, it loses the main advantages of the R-learner, since it is neither an unbiased estimator, nor neyman-orthogonal.


Paweł Morzywolek, Johan Decruyenaere, and Stijn Vansteelandt. On Weighted Orthogonal Learners for Heterogeneous Treatment Effects. arXiv:2303.12687, 2024.

Questions for authors
  • Could the authors clarify what limitation of the R-learner is addressed by the proposed structural drift? Specifically, the "partially linear represenation" is an exact identity and CATE error of the R-learner is second-order in nuisance errors.
  • Could the authors clarify what the regularized objective in Eq. 14 actually satisfies?
  • Could the authors address the counterexample above? Are there any additional assumptions necessary for the GRL to be unbiased?
9 comments

We thank the reviewer for the detailed comments. Theoretical motivation for the partially linear structure in PO v/s CATE estimation is addressed in G1 in our overall response, and GRL vs. R-learner in CATE estimation is addressed in G2 in our overall response.

What the regularized objective satisfies, and counterexample response: We thank the reviewer for the counterexample. It is correct: as we noted in G3 of our overall response, even with oracle nuisances, the penalty can bias the minimizer. In general, Eq. 14 satisfies the risk bound of Proposition 7 and the bias bound below, but it need not be unbiased or Neyman-orthogonal. Both properties hold under an additional assumption, which we have added to the revised paper (Assumption 12, Corollary 1, Appendix C.9):

(A) H\mathcal H contains an h0h_0 with h0(η0)=f0h_0(\eta_0)=f_0 and ∇lh0(η0)=∇eh0(η0)=0\nabla_l h_0(\eta_0)=\nabla_e h_0(\eta_0)=0 almost surely.

Let hˉ\bar h minimize Eq. 14 under oracle nuisances. Corollary 1 shows that:

  • in general, ∥hˉ(η0)−f0∥22≤Ω0∗:=inf⁡h∈H: h(η0)=f0Ω(h;η0)\Vert\bar h(\eta_0)-f_0\Vert_2^2\le\Omega_0^{\ast}:=\inf_{h\in\mathcal H:\ h(\eta_0)=f_0}\Omega(h;\eta_0), i.e., the bias is at most the smallest penalty among oracle-correct predictors;
  • under (A), Ω0∗=0\Omega_0^*=0, so hˉ(η0)=f0\bar h(\eta_0)=f_0 and the nuisance gradients of hˉ\bar h vanish; by Lemma 1, ψraw\psi_{\text{raw}} is then Neyman-orthogonal at hˉ\bar h;
  • under (A), the bound in Proposition 7 becomes O(∥δη∥2)O(\|\delta\eta\|^2).

The model that reviewer pointed out violates (A): since ∇lh=1\nabla_l h=1, Ω0∗≥E[(Y−l0)2]>0\Omega_0^*\ge\mathbb E[(Y-l_0)^2]>0, and its minimizer is τ0/2\tau_0/2, as the reviewer computed (Appendix C.9). We have also accordingly revised the claims about orthogonality and unbiasedness in the Abstract, Introduction, Section 5, & Conclusions.

Thank you for your rebuttal. I appreciate the effort of revising the paper. However, I my concerns have not been adequately addressed. I therefore maintain my original assessment.

1) Theoretical motivation: I appreciate that the authors now explicitly distinguish potential outcome and CATE estimation and acknowledge the second order nuisance property of the R-learner for CATE estimation.

However, this reinforces my original concern: the proposed "structural drift" is not a fundamental limitation of the R-learner for its CATE estimand. The first-order error in reconstructed in PO estimation is not a failure of the R learner since it is constructed for orthogonal CATE estimation. I therefore remain unconvinced by the central motivation for GRL.

2) Orthogonality: This is more severe. The authors acknowledge the counterexample and that their regularization can introduce bias even under oracle nuisance estimates.

However, the introduction of Assumption 12 is not satisfactory. The new assumption essentially states the existence of an oracle-correct predictor with vanishing nuisance gradients. This property is not given for the proposed model architectures, nor is it guaranteed by the regularization itself.

Even more importantly, Corollary 1 gives orthogonality of the residual moment at the oracle solution, but it does not demonstrate the second-order nuisance sensitivity of the CATE estimator that is obtained by minimizing the penalized objective.

The second-order excess-risk bound does not guarantee this property either. Hence, the proposed method sacrifices an important theoretical advantage of the original R-learner without providing an equivalent guarantee.


I can give another counterexample:

Constant XX, T∼Bernoulli(1/2)T\sim\mathrm{Bernoulli}(1/2), and Y=εY=\varepsilon, where ε\varepsilon is independent of TT, has mean zero and variance one. Thus, l0=0l_0=0, e0=1/2e_0=1/2, and the true CATE is τ0=0\tau_0=0.

Take, for example, the smooth second-stage model class

hθ(l,u)=l2+θ(u+l), u=T−e.h_\theta(l,u)=l^2+\theta(u+l),\, u=T-e.

At oracle nuisances and θ=0\theta=0, we have

h0(l0,T−e0)=0=f0, ∇lh0=∇eh0=0.h_0(l_0,T-e_0)=0=f_0, \, \nabla_l h_0=\nabla_e h_0=0.

Therefore, Assumption 12 is satisfied.

Take a small outcome nuisance error l^=δ\hat l=\delta and keep the propensity exact, e^=e0=1/2\hat e=e_0=1/2.

Substitute this into the population version of Eq. 14

LGRL−OP(θ)=E[(Y−hθ)2]+E[(∇lhθ)2(Y−l^)2]+E[(∇ehθ)2(T−e^)2].\mathcal L_{\mathrm{GRL-OP}}(\theta) =\mathbb E[(Y-h_\theta)^2]+\mathbb E[(\nabla_l h_\theta)^2(Y-\hat l)^2]+\mathbb E[(\nabla_e h_\theta)^2(T-\hat e)^2].

With ∇lhθ=2δ+θ\nabla_l h_\theta=2\delta+\theta and ∇ehθ=−θ\nabla_e h_\theta=-\theta, this simplifies to

=C_\delta+ \left(\frac32+2\delta^2\right)\theta^2 +(4\delta+6\delta^3)\theta,$$ where $C_\delta$ does not depend on $\theta$. The population minimizer is therefore $$\theta_\delta^* =-\frac{4\delta+6\delta^3}{3+4\delta^2} =-\frac43\delta+O(\delta^3).$$ Since the treatment contrast of $h_\theta$ equals $\theta$, the resulting CATE satisfies $$\hat\tau_{\mathrm{GRL}}-\tau_0=-\frac43\delta+O(\delta^3).$$ Thus, **the fitted GRL CATE has first-order sensitivity to nuisance estimation error although Assumption 12 is satisfied**. In contrast, the original population R-learner recovers $\tau_0=0$ **exactly** for every $\delta$ because the propensity is correctly specified. To be fair, this counterexample does not contradict the literal statement of Corollary 1 concerning orthogonality of the raw residual moment or its second-order excess risk bound. However, it demonstrates that these properties do **not** imply second-order nuisance sensitivity of the CATE estimator that one obtains by minimizing Eq. 14. Therefore, the new theoretical results do **not** imply that GRL retains the orthogonality advantage of the original R-learner. ____ Although the revisions improve the presentation and acknowledge important limitations, I the main theoretical motivation and claimed advantages of the regularized objective do **not** hold. I therefore maintain my original rating of **1 (Clear Rejection)**.
AuthorsOpenReview ↗

Technically correct comments but not targeted at the intended/stated contribution of GRL

We thank the reviewer for following up. However, we feel it important to point out that the reviewer's two major critiques -- while technically correct -- are based on a narrow interpretation of GRL's contribution that was not intended nor claimed in the paper.

First, GRL was never narrowly scoped or stated to address "the limitation of the R learner' in its CATE estimand". As laid out in Section 4, structural drift is a mathematical fact when plugging nuisance estimates into Robinson's residual-on-residual transformation on a partially linear outcome equation, and its existence will impact the subsequent PO and CATE estimation (albeit differently as we clarified in the revision). To reveal and mitigate this gap is the motivation of GRL --- to advance a seminal and powerful concept (i.e., Robinson's residual-on-residual transformation) beyond its original partially-linear formulation, so it can find wider utility beyond its current state-of-the-art. R-learner is a specific example of such a SoTA and an important baseline in the paper (for its intended CATE estimation as well as the PO derived from its CATE estimate), but GRL was never intended or claimed for the narrow scope of improving R-learner's CATE estimation.

Second, GRL was never stated to achieve exact orthogonality. On the contrary, throughout the paper, we have clearly stated that, to go beyond the partially linear equation (and the nice Neyman orthogonal estimating function it naturally lands on), we must now face the new challenge of "nuisance sensitivity" (Lemma 1) and propose the orthogonality regularized objective to suppress this sensitivity. We have acknowledged the loss of Neyman orthogonality in GRL throughout the paper: the newly added Assumption 12 is not to claim that it is always true, but to rigorously specify the condition under which the estimating function can be orthogonal versus not.

In summary, GRL was never intended or claimed by the narrow contribution of "a new orthogonal learner that improves the R-learner in CATE estimation", which seems to be what the reviewer's critiques are centered around. GRL was motivated to advance the classic Robinson's transformation beyond its original partially-linear formulation, to address its structural drift while mitigating the loss of orthogonality that arises in the process.

Reviewer_LdkNOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper starts from a target-error analysis of the classical R-learner for potential-outcome (PO) and CATE estimation. It shows that nuisance estimation errors induce a structural drift term that directly affects PO estimation and enters the finite-sample upper bound of CATE estimation. Under the R-learner's partially linear structure, this term is irreducible and cannot be controlled at the second stage. To address this, the authors propose the Generalized R-learner (GRL), which replaces the partially linear second-stage model with a nonlinear class and introduces an orthogonalization penalty that constrains the model's sensitivity to nuisance errors, thereby reducing the structural drift. Experiments on synthetic, semi-synthetic, and real-world datasets demonstrate the reliability and improved performance of the proposed method.

Critical strengths
  1. The paper is well-motivated and novel. Starting from a theoretical analysis showing that the R-learner's partially linear structure cannot control the structural drift induced by nuisance errors, the authors propose GRL, a concise and direct solution that replaces the partial-linearity with a nonlinear model class plus an orthogonalization penalty. The motivation is sound and clearly articulated.

  2. The theoretical results are thorough and sound. By combining the nonlinear model class with the orthogonalization penalty, GRL achieves better control of the PO estimation error and a tighter finite-sample upper bound for CATE estimation.

  3. The experimental evaluation is comprehensive. The method is tested on synthetic, semi-synthetic, and real-world benchmarks, with comparisons against a wide range of baselines from classical CATE estimators to recent state-of-the-art methods. The results are convincing.

  4. The presentation is clear and easy to follow. The paper effectively communicates the limitations of the existing R-learner, how GRL addresses them, and the resulting theoretical and empirical gains.

Critical weaknesses

The proposed improvements are specific to the R-learner's partially linear structure and do not directly carry over to other orthogonal estimators such as DR-learner. As a result, the applicability of GRL is limited to settings where the R-learner itself is preferred over alternatives. While the appendix discusses limitations of other orthogonal methods, the paper does not clearly delineate the specific scenarios (e.g., in terms of propensity score overlap or nuisance estimation quality) where GRL provides advantages over DR-learner or other orthogonal CATE estimators. In practice, R-learner and DR-learner often exhibit comparable performance when propensity scores are not extreme, so a clearer discussion of GRL's relative advantages would strengthen the paper.

Questions for authors

Please see Weaknesses.

3 comments

We thank the reviewer for the assessment and for prompting us to better delineate the advantage of GRL over other orthogonal learners such as the DR-learner. We considered the doubly robust property of the DR-learner, i.e., that the population DR-learner can recover the exact CATE when either the outcome or the propensity estimate is exact, in relation to the ability of the population R-learner to recover the exact CATE when the propensity estimate is exact (pointed out by Reviewer g8bU). Based on this, we hypothesized that the advantage of GRL may lie in scenarios where neither nuisance estimate is exact. To test this, we carried out a dedicated comparison of GRL, R-learner, and DR-learner under three scenarios of nuisance errors: nuisance error in l^(x)\hat{l}(x) only, e^(x)\hat{e}(x) only, and both. The detailed revisions and results are summarized in G2 of the overall response and Section 6.1, Figure 2, & Appendix F3 of the revised paper. In short, the findings were consistent with the theoretical analysis of the three learners: when either the propensity or the outcome nuisance estimate was exact, the DR-learner obtained the lowest PEHE, while the R-learner was only able to marginally outperform GRL when the propensity estimate was exact; in both of these settings, GRL had an advantage when the nuisance error in l^(x)\hat{l}(x) or e^(x)\hat{e}(x) was large. When neither nuisance was exact, GRL outperformed both the R-learner and DR-learner even at relatively small nuisance errors, and remained consistently robust as the nuisance error increased. These findings helped identify concrete use scenarios for GRL, which we believe strengthened our work and we thank the reviewer again for encouraging us to probe in this direction.

As the discussion period is drawing to a close, we would like to follow up with the reviewer to see if there are any additional questions we can address in the revised paper. Thank you.

Thank you for the response. I will keep my score.

Comments on the paper

3 comments

General Response (1/2)

We thank all reviewers for their constructive feedback. We have uploaded a revised version, with main changes marked in blue and summarized below.

G1. Clarification of GRL's contribution in PO vs. CATE estimation (Reviewer g8bU)

  • In response to Reviewer g8bU’s first major critique, we would like to first clarify a possible misunderstanding: we completely agree with the reviewer and already specified in Section 3 (below Equation (3)) of the original manuscript that the partially linear outcome equation used in the R-learner, E[Y∣X,T]=l0(X)+(T−e0(X))τ0(X)\mathbb{E}[Y|X,T] = l_0(X) + (T-e_0(X))\tau_0(X), is exact under binary interventions and oracle nuisances regardless whether the true outcome equation is linear or nonlinear. The problem arises when we plug estimated nuisances l^\hat{l} and e^\hat{e} into the same partially linear structure l^(X)+(T−e^(X))τ(X)\hat{l}(X) + (T-\hat{e}(X))\tau(X) as the plug-in target: when l^≠l0\hat{l} \neq l_0 and e^≠e0\hat{e} \neq e_0, the nuisance errors induce a structural distortion between the oracle and optimal plug-in outcome target; the use of partially linear outcome equation for the latter cannot absorb this drift, which will affect both outcome and CATE estimation. We had originally added the brief clarification in Section 3 in anticipation of similar confusions; in light of Reviewer g8bU's question, we have added further elaborations on this distinction in Abstract and Section 4 when contrasting the oracle (Equation 5) with plug-in target (Equation 6).
  • Structural drift for PO vs. CATE estimation: We do agree that this nuisance-induced structural drift impacts PO and CATE estimation differently: the nuisance errors enter PO estimation error at first order, and CATE error at second order. We thank Reviewer g8bU for the suggestion to better delineate this. In response, we have revised the manuscript throughout to clarify this distinction (Contribution I in Introduction and Sections 4.1-4.2 for PO and CATE analysis separately), specified whether the discussion is about the PO or CATE whenever appropriate, and made sure that the structural drift is discussed in the concrete context of using the partially linear equation as plug-in outcome target, rather than loosely referring to it as a limitation for the R-learner.

G2. When should GRL be preferred over R- and DR-learner (Reviewer g8bU and LdkN)
We further thank Reviewer g8bU for pointing out that the population R-learner recovers the exact CATE when e^=e0\hat{e} = e_0. We relate this to Reviewer LdkN's suggestion for us to demonstrate GRL's advantage over DR-leaner, which recovers the exact CATE when either e^=e0\hat{e} = e_0 or l^=l0\hat{l} = l_0 at the population level. We thus took this opportunity to strengthen our manuscript by adding a dedicated analysis of the impact of nuisance errors on the R-learner, DR-learner, and GRL under three distinct conditions: nuisance error in (1) l^(x)\hat{l}(x) only; (2) e^(x)\hat{e}(x) only; and (3) both. Our empirical results (Section 6.1, Fig. 2, & Appendix F.3) support the theoretical analysis for the three learners. When e^(x)=e0(x)\hat{e}(x)=e_0(x) was exact and nuisance error in l^(x)\hat{l}(x) was not large, DR-learner produced the lowest PEHE while R-learner's PEHE was marginally lower than GRL; when l^(x)=l0(x)\hat{l}(x) = l_0(x), the DR-learner retained the best PEHE (demonstrating its doublely robustness) when error in e^(x)\hat{e}(x) was not large, whereas the R-learner only outperformed GRL at e^(x)=e0(x)\hat{e}(x)=e_0(x) and deteriorated significantly after. In both cases, PEHE of the R- and DR-learner deteriorated much faster than GRL as the nuisance error increased, such that GRL provided the lowest PEHE at larger nuisance errors. When neither of the nuisance was exact, GRL outperformed both R- and DR-learners even at small nuisance errors (e.g., nuisance outcome error ~ 10% and propensity error ~40%). Similar trends were observed for the PO MSE when comparing GRL and R-learner (PO not available from the DR learner). These empirical evidence provided concrete evidence for when GRL can be preferable: when one of the nuisance errors is large (e.g., over 40% error in nuisance outcome), when neither nuisance is exact even when the error may be small (e.g., ~10% error in nuisance outcome), or when the PO estimate is of interest.

G3. Regularized objective. (Reviewer g8bU and tLXP)
We have added a new assumption and corollary (Section 5 & Appendix C9) to specify when the orthogonality regularized objective is unbiased and Neyman-orthogonal, with a second-order risk bound. We have also added related discussion of limitation in Section 7.

General Response (2/2)

G4. Addition of recent baselines (Reviewer tLXP). We have added all three suggested baselines to the discussion of related works (Section 2) and experiments (Section 6, 6.1, & Tables 1-2): HyperITE [3] with TARNet backbone, H-learner [1] with X-learner's pseudo-outcome, and PairNet [2]. On synthetic data, PairNet achieved the lowest PEHE for all values of λ\lambda, but had considerably higher PO MSE as its pair loss promotes positive correlation between treatment arms which cancel in CATE, but not in potential outcomes. The H-learner achieved lower PEHE than GRL at only higher values of λ\lambda (λ≥0.8\lambda \ge 0.8), but had higher PO MSE that also fluctuated across different values of λ\lambda. HyperITE outperformed TARNet for λ≥0.2\lambda \ge 0.2, but remained behind GRL in both PEHE and PO MSE. On real benchmarks, our earlier observations remain unchanged: GRL had the lowest PEHE on IHDP, while differences on Twins and Jobs remained small and not statistically significant across top-performing methods; while the newly-added PairNet achieved the lowest policy risk on Jobs, its performance on Twins was sub-optimal and GRL remained the only method that was consistently among the top-performers across datasets.

References:

[1] Liang, Zhongyuan, Lars van der Laan, and Ahmed Alaa. "Hybrid Meta-learners for Estimating Heterogeneous Treatment Effects."

[2] Nagalapatti, Lokesh, et al. "PairNet: Training with observed pairs to estimate individual treatment effect."

[3] Chauhan, Vinod Kumar, et al. "Dynamic inter-treatment information sharing for individualized treatment effects estimation."

We'd like to thank Reviewer g8bU for the continued discussions and Reviewer tLXP for the appreciation of the revised manuscript. In response to the feedback from both reviewers, we have made further improvements to the manuscript as summarized below.

O1. More precise articulation of how structural drift enters CATE error

To further improve the delineation of GRL's contribution in PO vs. CATE estimation, we now more precisely frame how exactly the structural drift impacts CATE error. As shown in Equation 32, CATE error is proportional to Δe(Δl−Δeτ0)\Delta e (\Delta l - \Delta e \tau_0), where Δl−Δeτ0\Delta l - \Delta e \tau_0 is the leading term of the function-level structural distortion term defined in Proposition 3. We therefore made minor edits to Proposition 3 to explicitly define this term and, instead of generally stating that the nuisance error enters the CATE estimation at second order as in our last revision, we made minor edits to precisely state that the structural drift enters the CATE error after modulation by Δe\Delta e, and that it does not bias CATE error only in the special case when Δe=0\Delta e = 0. These edits are reflected in Introduction -- Contribution I and Section 4.2. We hope that this both helps delineate when the structural drift does bias the CATE estimation (whenever Δe≠0\Delta e \neq 0), and clarify why the structural integrity of Robinson's residualization -- the focus of GRL -- is foundational to the derivation of its downstream target estimators (including both PO and CATE estimation).

To strengthen the empirical evidence, in Section 6.1 where we tested nuisance error under 3 different scenarios, in Scenarios 2 and 3 where Δe≠0\Delta e \neq 0, we refined the level of nuisance error tested for e^\hat{e} in between 0%-33% relative error. The results showed that, when Δl=0\Delta l=0, R-learner started to underperform GRL in PEHE at $\Delta e \ge $ 23%; when Δl≠0\Delta l \neq 0, R-learner started to underperform GRL at a smaller error in e^(X)\hat{e}(X), e.g., 19%19\% error in e^(X)\hat{e}(X) with 9%9\% error in l^(X)\hat l(X). Figure 2 and discussion in Section 6.1 is lightly edited accordingly.

O2. Y(T,X)=l^(X)+(T−e^(X))τ(X)+a(X)Y(T,X) = \hat l(X) + (T-\hat e(X))\tau(X) + a(X) as a specific form to generalize Robinson's residualisation approach

In response to Reviewer g8bU's comment, we also considered the special case where the plug-in target outcome takes an augmented partially-linear (aPL) structure Y(T,X)=l^(X)+(T−e^(X))τ(X)+a(X)Y(T,X) = \hat l(X) + (T-\hat e(X))\tau(X) + a(X), where the structural drift is explicitly resolved in a correction function a(X)a(X). This variant is theoretically tighter but imposes a factorization of two separate functions at the second stage, a(X)a(X) and τ(X)\tau(X), whose joint optimization is structurally ill-conditioned as they share the oracle τ0(X)\tau_0(X) as a latent component while the complexity of a(X)a(X) grows with propensity errors and nonlinearity of the oracle outcomes. Revisions are made to Section 6.1 -- GRL ablation and complete results are included in Appendix F4 across all datasets, which demonstrated a tradeoff that favored aPL when the oracle outcome structure was approximately additive and nuisance errors were moderate, and favored GRL when either condition broke down.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.