acceptodds
Under review as a conference paper at ICLR 2027

NEAR: Neighborhood Effect-Aware Regularization for Two-Stage CATE Estimation

Abstract

Estimating the conditional average treatment effect (CATE) from observational data is a central task in causal inference. A common approach is two-stage CATE learning, which first estimates nuisance functions and then uses them to fit a target CATE model. However, the resulting second-stage CATE targets or losses depend on estimated nuisance functions and can be noisy. Existing methods use regularization and information sharing to stabilize second-stage estimation, yet it remains unclear how noisy first-stage CATE estimates should guide the selection of observations to regularize together. We propose Neighborhood Effect-Aware Regularization (NEAR), a modular second-stage regularizer that uses first-stage plug-in CATE estimates to construct a fixed graph and regularizes CATE predictions over pairs with similar plug-in CATEs. We show how plug-in CATE estimation errors control true CATE variation along the selected pairs and derive finite-sample CATE risk bounds for squared-loss second-stage regression using a stabilized version of the doubly robust pseudo-outcome. Across six two-stage learners and three benchmarks, NEAR lowers mean CATE estimation error. Graph support comparisons and perturbation diagnostics further support using plug-in CATE similarity to select pairs for regularization.

Primary area:
causal reasoning
Keywords:
Causal inference

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_X2CQOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes NEAR, a modular regularizer that constructs a fixed graph from first-stage plug-in CATE similarity and penalizes differences between connected second-stage CATE predictions. The authors link plug-in estimation error to true CATE variation along selected edges, provide risk and bias–variance analyses, and demonstrate consistent improvements across six learners and three benchmarks. Remaining concerns include novelty relative to existing effect-aware pairing methods, the applicability of the theoretical analysis to practical training, and the incremental benefits over strong plug-in baselines.

Critical strengths

(1) The work motivatedly uses plug-in CATE similarity rather than covariate proximity alone, particularly when treatment effects change sharply between observations with similar covariates.

(2) NEAR constructs a fixed mutual top-K graph before second-stage training and applies pairwise regularization only to the final CATE predictions, leaving the learner-specific base loss unchanged. Moreover, plug-in CATE estimates define the neighborhood structure rather than serve as numerical anchors for the final predictions.

(3) The theoretical analysis and empirical results connect plug-in estimation error, true CATE variation along selected edges, and the cost of smoothing, providing an interpretable framework for understanding the method’s behavior.

Critical weaknesses

(1) Although the paper discusses related work on Pairwise and local structure in CATE learning and Graph and fusion regularization, and effect-aware pairing, the key methodological distinctions and incremental contributions could be articulated more clearly. A specific comparison with methods (such as HERMES and LITE) would help clarify the work's novelty.

(2) Although NEAR achieves lower mean estimation error than directly using the first-stage plug-in CATE estimates, it also introduces additional computational overhead that warrants further analysis.

Questions for authors

similar to Weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer X2CQ (1/2)

Thank you for recognizing our use of effect similarity and the connection between plug-in errors and the cost of smoothing! We clarify the distinctions from HERMES and LITE and analyze the additional computational cost below.

[W1] Distinctions from HERMES and LITE

NEAR constructs fixed neighborhoods from first-stage effect estimates and directly regularizes final conditional average treatment effect (CATE) predictions. The differences are how we select pairs and what we regularize.

1. HERMES dynamically forms pairs from current effect predictions during training and regularizes their representations, while NEAR fixes its pairs before the second stage and directly regularizes CATE predictions.

HERMES (Zaccagnino et al., 2026) also uses estimated effect similarity. Its Siamese network encourages similar representations for similar effects and separates representations for dissimilar effects. NEAR determines its neighborhoods from existing first-stage estimates before second-stage training and holds them fixed throughout that stage. Its regularizer penalizes squared differences between final effect predictions over connected pairs.

2. LITE separately regularizes treatment and control outcome predictions, while NEAR directly regularizes CATE predictions without requiring similar treatment and control outcome levels.

The graph in LITE (Cohen et al.) is based on learned representation distances and changes during training. NEAR constructs a fixed graph from differences between scalar plug-in CATE estimates.

3. NEAR reuses existing first-stage estimates, retains the original learner's base loss and prediction model, and analyzes how first-stage estimation errors affect second-stage risk.

This design requires no additional model for neighbor selection. Our analysis further connects plug-in errors to true CATE variation along selected pairs and to second-stage CATE risk.

4. In the additional comparisons, the doubly robust learner with NEAR (DR-NEAR) has the lowest mean error on IHDP and ACIC, while LITE has the lowest mean error on Synthetic.

Table A compares the submitted DR-NEAR results with results for these methods and Causal Isotonic Calibration (CIC). For CIC, we use Algorithm 3 of van der Laan et al. (2023). All error tables report outcome-standardized root PEHE, consistent with the paper, as mean ± standard deviation. Lower is better.

Table A. Comparison with existing methods.

DatasetDR-NEARLITEHERMESCIC
Synthetic0.47 ± 0.070.39 ± 0.060.98 ± 0.060.62 ± 0.05
IHDP0.49 ± 0.220.55 ± 0.190.66 ± 0.280.69 ± 0.29
ACIC0.47 ± 0.220.51 ± 0.230.74 ± 0.341.89 ± 6.00

The ACIC row includes all 77 settings. Excluding the three settings with extreme CIC errors gives 0.90 ± 0.44 for CIC on the remaining 74 settings.

Thus, NEAR selects fixed neighborhoods from existing first-stage effect estimates and regularizes final CATE predictions while retaining the original learner's base loss and prediction model.

AuthorsOpenReview ↗

Response to Reviewer X2CQ (2/2)

[W2] Computational cost relative to Direct

In the new runtime comparison, most of DR-NEAR's additional time over direct plug-in prediction (Direct) comes from the propensity fitting and second-stage learning required by the DR-learner without graph regularization (None). Adding NEAR to None increases total time by 20.9% to 23.4% across these four runs.

1. NEAR builds the graph once and computes the regularizer during training. Prediction does not need the graph.

Let nn be the number of training observations and KK the neighbor limit. The current implementation constructs the graph once using O(n2)O(n^2) candidate pair evaluations. For large datasets, graph construction can be further optimized. Since similarity depends only on differences between scalar plug-in CATE estimates, sorting these estimates before radius filtering and neighbor selection can reduce the complexity of graph construction to O(nlog⁡n)O(n\log n) for fixed KK. Table B reports measured times for the current implementation.

Mutual neighbor selection retains at most nK/2nK/2 edges, requiring O(nK)O(nK) memory to store the retained edges and weights. Algorithm 2 evaluates the regularizer on sampled edges during training, with the graph held fixed.

2. At every sample size tested, the time increase from Direct to None exceeds the increase from None to DR-NEAR.

We conducted the new runtime comparison on Synthetic, with one run at each sample size. Table B includes the fitting and prediction steps required by each method.

Table B. Measured computation times for Direct, None, and DR-NEAR, in seconds.

Total sample sizeDirect totalNone totalDR-NEAR graph constructionDR-NEAR second stageDR-NEAR total
1,0003.0911.170.028.4313.52
2,0006.3122.450.0817.0027.15
5,00015.9757.790.5143.4069.88
10,00032.76117.842.3389.65145.37

DR-NEAR takes 4.30 to 4.44 times as long as Direct. At a sample size of 10,000, None adds 85.08 seconds over Direct, while DR-NEAR adds 27.53 seconds over None.

Thus, NEAR adds graph construction and regularizer evaluations to existing two-stage learning. Direct is faster, but Table 4 of the revised manuscript (Table 3 of the original submission) reports higher mean estimation error for Direct.

We again thank you for these valuable suggestions! In the revised manuscript, we have clarified the methodological distinctions in Section 2, added the method comparison in Table 2, and explained graph construction complexity and training costs in Section 6.3 and Appendix C.9. Your comments really helped us clarify NEAR's contribution and computational cost. We hope this addresses your concerns.

AuthorsOpenReview ↗

References

[1] Cohen, D., Talmon, R., & Shaham, U. (2025). Revisiting covariate and hypothesis roles in ITE estimation: A new approach using Laplacian regularization. OpenReview manuscript.

[2] Zaccagnino, R., Benevento, G., Malandrino, D., Telesca, D., & Ture, A. (2026). HERMES: Heterogeneous effects representation with matched embeddings using Siamese networks. Transactions on Machine Learning Research.

[3] van der Laan, L., Ulloa-Perez, E., Carone, M., & Luedtke, A. (2023). Causal isotonic calibration for heterogeneous treatment effects. ICML.

Reviewer_HPj2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes NEAR, a modular regularizer for two-stage CATE estimation. It constructs a fixed graph from first-stage plug-in estimates using radius filtering and mutual top-K selection, then penalizes differences between second-stage predictions along selected edges. The analysis links plug-in errors to true CATE variation on these edges and derives a finite-sample bound on CATE function risk.

Critical strengths

The paper addresses an important issue: covariate similarity does not necessarily imply similar treatment effects. It uses first-stage CATE estimates to guide second-stage regularization and includes ablation studies covering several key components, including graph construction, edge weighting, and the use of first-stage estimates.

Critical weaknesses
  1. Unclear theoretical presentation. Terms such as “weighted-degree factor,” “reference noise,” and “comparator” are insufficiently explained or clearly linked to their mathematical definitions, making the theoretical analysis difficult to follow.

  2. Incomplete experimental comparisons. Main comparisons and ablation studies should be more clearly separated. Broader comparisons with existing methods are needed. The experimental settings should be made more consistent. For example, Tables 2 and 4 should also include results on the Synthetic benchmark.

  3. Insufficiently controlled sensitivity analyses. Panels (a)–(c) use different experimental conditions. A complementary analysis should use the same dataset and data splits, with a common reference parameter configuration, varying one parameter at a time.

Questions for authors
  1. Could the authors explain why the error fluctuates as K varies at \lambda=10 in Figure 2(b)?
  2. Figure 2(c) shows little change in performance as the radius varies from 0.01 to infinity. Could the authors clarify how these results support the mechanism illustrated in Figure 1?
5 comments
AuthorsOpenReview ↗

Response to Reviewer HPj2 (1/4)

Thank you for recognizing the motivation for using effect similarity and our ablation studies! We address your concerns about the theoretical presentation, method comparisons, sensitivity analysis, and the connection between Figures 1 and 2 below.

[W1] Meaning of the theoretical terms

The weighted-degree factor quantifies how concentrated edge weights affect the error bound. Reference noise is the random error that remains with true outcome regressions and propensity scores. The comparator provides a reference for the risk bound that includes both prediction error and graph regularization cost.

1. The weighted-degree factor captures the concentration of edge weights and enters the upper bound on the weighted sum of squared estimation error differences along selected edges.

Let nn be the number of training observations and wijw_{ij} the normalized edge weight between observations ii and jj. The maximum weighted degree is

dmax⁡w=max⁡i∑jwij,d_{\max}^w=\max_i\sum_j w_{ij},

the largest total edge weight incident to one observation. In Equation (13), the weighted-degree factor 2ndmax⁡w2n d_{\max}^w multiplies the mean squared deviation of first-stage effect estimation errors from their sample mean, bounding the weighted sum of squared error differences along selected edges.

The neighbor limit KK bounds the number of connections, while the weighted degree further accounts for the concentration of normalized edge weights across observations.

2. Reference noise measures the random error of the pseudo-outcome relative to the true effect when the true outcome regressions and propensity score are used.

Reference noise is Zi0−τ(Xi)Z_i^0-\tau(X_i), where τ(Xi)\tau(X_i) is the true conditional average treatment effect (CATE) for observation ii and Zi0Z_i^0 is the stabilized doubly robust (DR) pseudo-outcome evaluated at the true outcome regressions and propensity score. The pseudo-outcome is the response used in second-stage regression.

Estimating the outcome regressions and propensity score introduces a further difference, Z^i−Zi0\widehat Z_i-Z_i^0, where Z^i\widehat Z_i is the fitted pseudo-outcome. Theorem 1 measures its root mean square over training observations by ζn\zeta_n, separating nuisance estimation error from reference noise.

3. The comparator provides a reference for the risk bound that includes both CATE prediction error and graph regularization cost.

A comparator gg belongs to the same fixed function class F\mathcal F as the fitted function. The comparator term in Theorem 1 is

inf⁡g∈F{R(g)+λΩ(g)}.\inf_{g\in\mathcal F}\{\mathcal R(g)+\lambda\Omega(g)\}.

Here, R(g)\mathcal R(g) is the mean squared CATE prediction error on new observations from the same population, Ω(g)\Omega(g) is the weighted sum of squared prediction differences along selected edges, and λ\lambda is the regularization strength. This term accounts for both how well gg approximates the true CATE and the cost of smoothing it over the graph.

Thus, the weighted-degree factor controls the upper bound on the graph term involving estimation errors. Reference noise is treated separately from nuisance estimation error, and the comparator accounts for both prediction error and the cost of graph smoothing.

AuthorsOpenReview ↗

Response to Reviewer HPj2 (2/4)

[W2] Broader method comparisons and Synthetic results

Following your suggestions, we have expanded the method comparison and added Synthetic results to the regularizer and graph construction ablations, with method comparisons and ablations presented separately in the revised manuscript.

The updated ablations appear in Tables 3 and 5 of the revised manuscript (Tables 2 and 4 of the original submission). All error tables report outcome-standardized root PEHE as mean ± standard deviation. Lower is better.

1. Among the DR-learner with NEAR (DR-NEAR), LITE, HERMES, and CIC, DR-NEAR has the lowest mean error on IHDP and ACIC, while LITE has the lowest mean error on Synthetic.

Table A reports the submitted DR-NEAR results alongside results for LITE (Cohen et al.), HERMES (Zaccagnino et al., 2026), and Causal Isotonic Calibration (CIC). For CIC, we use Algorithm 3 of van der Laan et al. (2023).

Table A. Comparison of prediction errors with existing methods.

DatasetDR-NEARLITEHERMESCIC
Synthetic0.47 ± 0.070.39 ± 0.060.98 ± 0.060.62 ± 0.05
IHDP0.49 ± 0.220.55 ± 0.190.66 ± 0.280.69 ± 0.29
ACIC0.47 ± 0.220.51 ± 0.230.74 ± 0.341.89 ± 6.00

The ACIC row includes all 77 settings. Excluding the three settings with extreme CIC errors gives 0.90 ± 0.44 for CIC on the remaining 74 settings.

2. With the added Synthetic results, NEAR has the lowest mean error among the compared regularizers on all three datasets.

We conducted 10 additional Synthetic repetitions with the same DR-learner. Table B presents the updated regularization comparison, also reported in Table 3 of the revised manuscript.

Table B. Comparison of regularization strategies.

RegularizerIHDPACICSynthetic
None0.69 ± 0.340.60 ± 0.200.93 ± 0.14
Constant0.65 ± 0.330.58 ± 0.210.81 ± 0.15
dOAR0.61 ± 0.300.55 ± 0.210.69 ± 0.10
NEAR0.49 ± 0.220.47 ± 0.220.47 ± 0.07

None denotes no added regularization, Constant denotes regularization at a fixed level, and dOAR denotes debiased Overlap-Adaptive Regularization.

3. NEAR with uniform edge weights (NEAR-U) has lower mean error than the other compared graph constructions on all three datasets.

Table C reports the graph construction comparison, including results from 10 Synthetic repetitions, as in Table 5 of the revised manuscript. All graphs use uniform edge weights, with regularization strength selected separately by validation.

Table C. Comparison of graph constructions with the doubly robust learner.

DatasetNoneCovariateRandomShuffledNEAR-U
IHDP0.69 ± 0.340.74 ± 0.130.84 ± 0.140.85 ± 0.140.51 ± 0.26
ACIC0.60 ± 0.200.62 ± 0.230.66 ± 0.280.66 ± 0.290.49 ± 0.21
Synthetic0.93 ± 0.140.55 ± 0.090.65 ± 0.070.62 ± 0.070.50 ± 0.12

Covariate constructs edges based on covariate similarity, Random uses random pairs, and Shuffled constructs the graph after shuffling the first-stage effect estimates.

Thus, the broader method comparison supports DR-NEAR's lower mean error on IHDP and ACIC, while the ablations support NEAR's regularization and pair selection based on effect similarity across all three datasets.

AuthorsOpenReview ↗

Response to Reviewer HPj2 (3/4)

[W3, Q1] Sensitivity analysis under common conditions and error fluctuations with KK

We have added sensitivity analyses under common experimental conditions. Increasing KK does not always lower mean error. This is consistent with a fixed regularization strength λ\lambda acting on different pairs and weights as KK changes.

1. The additional analyses vary one parameter at a time on the same IHDP splits with a common reference configuration, providing the controlled comparison you requested.

The analyses use the same 10 IHDP splits, shared first-stage nuisance predictions, and the reference configuration (λ,K,ρ)=(10,8,1.7941)(\lambda,K,\rho)=(10,8,1.7941). Here, ρ\rho is the radius for filtering candidate pairs. We retain the reference radius from Figure 2(b), ρ≈1.794\rho\approx1.794, which corresponds to a Gaussian similarity threshold of 0.2.

Tables D and E report the results for KK and λ\lambda. The results for ρ\rho appear in Q2 under the same conditions.

Table D. Varying KK with λ=10\lambda=10 and ρ=1.7941\rho=1.7941 fixed.

KKError
10.68 ± 0.24
20.66 ± 0.23
40.66 ± 0.23
80.66 ± 0.22
160.66 ± 0.21
320.66 ± 0.20
640.67 ± 0.17
1280.68 ± 0.14
6710.78 ± 0.08

Table E. Varying λ\lambda with K=8K=8 and ρ=1.7941\rho=1.7941 fixed.

λ\lambdaError
0.10.79 ± 0.29
10.68 ± 0.25
100.66 ± 0.22
1000.66 ± 0.18
10000.83 ± 0.12
100000.92 ± 0.22

2. Fixing λ\lambda does not fix the regularization applied to each pair, because changing KK changes both the selected pairs and their normalized weights.

Although the weights sum to 1 for every nonempty graph, their distribution across pairs changes with KK. In Figure 2(b), the curves for different KK reach their lowest error at different values of λ\lambda. Thus, λ=10\lambda=10 does not minimize error for every curve.

Thus, the nonmonotonic changes in mean error observed under common conditions are consistent with the changing pairs and weights at fixed λ\lambda.

Reviewer_NnypOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Neighborhood Effect-Aware Regularization, a graph-based regularization method for two-stage CATE estimation. NEAR constructs a graph connecting observations with similar first-stage plug-in CATE estimates and encourages similar second-stage CATE predictions among connected observations. The authors provide theoretical results relating plug-in estimation errors, graph structure, and CATE estimation risk, along with a bias–variance analysis. Experiments on synthetic, IHDP, and ACIC datasets demonstrate improved CATE estimation accuracy across six different two-stage learners.

Critical strengths
  • The proposed method is simple, intuitive, and easy to implement. The intuition behind encouraging similar CATE predictions for observations with similar estimated treatment effects is clear and well motivated.
  • The authors justify using relative plug-in CATE information and establish translation invariance, partially relaxing the requirements on nuisance estimation accuracy.
  • The experiments are well organized, with comprehensive comparisons and ablation studies that effectively demonstrate the advantages of NEAR.
  • The authors provide careful theoretical analysis. Theorem 1/2 provides finite-sample guarantees for data-dependent graphs and nuisance estimates constructed from the same sample, without requiring sample splitting. The analysis also carefully handles several boundary cases.
Critical weaknesses
  • Limited novelty. The proposed method primarily adapts standard graph Laplacian regularization to CATE estimation. Although the implementation and analysis are well executed, the methodological contribution appears relatively incremental.

  • Strong dependence on nuisance estimation quality. Translation invariance only addresses common additive errors, however, nuisance estimation failures typically involve missing important treatment effect heterogeneity. Such errors can cause inappropriate neighborhood selection and introduce additional bias. Moreover, if the plug-in CATE estimates are accurate enough to construct reliable neighborhoods, direct plug-in estimation may already perform well. This raises questions about when NEAR offers substantial additional benefits.

Questions for authors
  • How does NEAR fundamentally differ from existing approaches such as Causal Isotonic Calibration for Heterogeneous Treatment Effects (van der Laan et al., 2023)? Both methods leverage estimated treatment effects to guide information sharing and improve CATE predictions. A more detailed discussion of the conceptual differences and a direct empirical comparison would help clarify the novelty and advantages of NEAR.

  • The main risk bounds apply to arbitrary nonnegative graphs. Could the authors further clarify what the theoretical results establish specifically about NEAR's graph construction, beyond standard graph regularization?

  • Could the authors investigate scenarios where the first-stage estimates systematically miss certain effect modifiers or distort the relative ordering of treatment effects? Such experiments would help establish when NEAR remains beneficial and when its regularization may introduce additional bias.

Reference:

Causal isotonic calibration for heterogeneous treatment effects. van der Laan et al. ICML 2023

5 comments
AuthorsOpenReview ↗

Response to Reviewer Nnyp (1/3)

Thank you for recognizing NEAR's simple design, its use of relative plug-in information, and the theoretical and experimental analyses! We address your concerns about the methodological contribution, the theory for graph construction, and dependence on first-stage estimation quality below.

[W1, Q1] Methodological contribution and difference from CIC

NEAR's contribution is to select fixed neighborhoods for second-stage conditional average treatment effect (CATE) regularization using existing first-stage estimates and to analyze how estimation errors affect true effect differences along selected pairs and second-stage risk.

1. NEAR uses first-stage effect similarity to select which second-stage CATE predictions to regularize together, while retaining the learner's base loss and prediction model.

NEAR takes the difference between first-stage treatment and control outcome predictions as the CATE estimate and connects observations with similar estimated effects. During the second stage, it penalizes squared differences between connected CATE predictions and keeps the graph fixed. Neighbor selection requires no additional model. We evaluate this design across six two-stage learners and three datasets.

2. Causal Isotonic Calibration (CIC) learns a nondecreasing mapping of initial predictions in its calibration step, while NEAR's second-stage predictions need not preserve the initial ordering.

CIC's calibration step takes the initial effect predictions as input, while NEAR's second-stage model takes covariates as input. In NEAR's graph regularizer, the initial estimates determine the neighborhoods and edge weights. The penalty acts on prediction differences between connected observations.

3. In the additional comparison, the doubly robust learner with NEAR (DR-NEAR) has lower mean error than CIC on all three datasets.

Following your suggestion, Table A compares the submitted DR-NEAR results with CIC. We use Algorithm 3 of van der Laan et al. (2023) for CIC. All error tables report outcome-standardized root PEHE as mean ± standard deviation. Lower is better.

Table A. Comparison of prediction errors with CIC.

DatasetDR-NEARCIC
Synthetic0.47 ± 0.070.62 ± 0.05
IHDP0.49 ± 0.220.69 ± 0.29
ACIC0.47 ± 0.221.89 ± 6.00

The ACIC row includes all 77 settings. Excluding the three settings with extreme CIC errors gives 0.90 ± 0.44 for CIC on the remaining 74 settings.

Thus, NEAR contributes pair selection and its error analysis for second-stage CATE regularization. DR-NEAR also has lower mean error than CIC on all three datasets.

AuthorsOpenReview ↗

Response to Reviewer Nnyp (2/3)

[Q2] What the theory establishes about NEAR's graph construction

For second-stage regression with squared loss and a stabilized DR pseudo-outcome, the main risk bound provides a general framework for nonnegative graphs. The analysis of NEAR's graph construction further shows how first-stage errors affect true effect differences along selected edges and how these differences enter the regularization cost in the second-stage risk bound.

1. A finite radius bounds estimated effect differences along selected edges, while differences in estimation errors also control the true effect differences.

Let τi\tau_i denote the true CATE for observation ii and δi\delta_i its first-stage effect estimation error, defined as the estimate minus the true effect. For finite selection radius ρ\rho, Lemma 1 gives

\leq \rho\sigma_\tau+\sqrt{\Omega_{K,\rho}(\delta)}.$$ Here, $\sigma_\tau$ is the effect scale used for graph construction, $K$ is the neighbor limit, and $\Omega_{K,\rho}$ is the normalized weighted sum of squared differences along selected edges. Thus, the bound on true effect differences combines the radius constraint with differences in estimation errors along selected edges. A common additive error cancels from these differences. **2. NEAR's pair selection bounds the graph regularization cost of a candidate effect function, which enters the second-stage risk bound.** The risk bound in Theorem 1 includes the regularization cost $\lambda\Omega_{K,\rho}(g)$ incurred by a candidate CATE function $g$, where $\lambda$ is the regularization strength. Combining Lemma 1 with Equation (16) gives $$\sqrt{\Omega_{K,\rho}(g)} \leq \rho\sigma_\tau +\sqrt{\Omega_{K,\rho}(\delta)} +\sqrt{2n d_{\max}^w Q_n(g)}.$$ Here, $n$ is the number of training observations, $Q_n(g)$ is the mean squared difference between $g$ and the true CATE on those observations, and $d_{\max}^w$ is the maximum total edge weight incident to a single observation. NEAR's mutual neighbor selection connects each observation to at most $K$ neighbors. Together with the selection radius and the number of retained edges, this yields a further upper bound on $d_{\max}^w$. **Thus, NEAR's radius, estimation errors along selected edges, and graph structure jointly bound the regularization cost. Equation (16) connects these graph properties to the main risk bound.**
AuthorsOpenReview ↗

Response to Reviewer Nnyp (3/3)

[W2, Q3] Benefit over Direct and failures in first-stage estimation

Reliable neighborhoods do not mean that direct plug-in prediction (Direct) is already accurate. NEAR can still benefit from relative effect information. Our new experiments also show when errors in that information remove the benefit.

1. DR-NEAR has lower mean error than Direct on all three datasets, and its second-stage predictions are not required to match the initial effect estimates.

NEAR learns the final CATE through the second-stage base loss and graph regularizer. The graph regularizer does not require predictions to be close to the initial effect estimates. Table B compares DR-NEAR and Direct in the submitted experiments.

Table B. Comparison with Direct.

DatasetDirectDR-NEAR
Synthetic0.52 ± 0.050.47 ± 0.07
IHDP0.53 ± 0.270.49 ± 0.22
ACIC0.50 ± 0.220.47 ± 0.22

Table 8 of the revised manuscript (Table 7 of the original submission) further shows that DR-NEAR has lower error than Direct on 70 of 100 IHDP splits and 65 of 77 ACIC settings.

2. In the additional experiments, omitting an effect modifier only from the model used for graph construction eliminates DR-NEAR's advantage in mean error over Direct.

Following your suggestion, we conducted 10 additional Synthetic repetitions. The true CATE is −1-1 when X1<0X_1<0 and 22 when X1≥0X_1\geq0, so X1X_1 determines effect heterogeneity in this setting.

Table C examines omitting X1X_1 only from the model used for graph construction and omitting it from both the first-stage outcome regressions and the propensity network. The second stage retains all covariates in both settings. The first setting replaces only the first-stage estimates used for graph construction, so Direct reuses its original predictions and its error remains unchanged.

Table C. Effect of omitting X1X_1.

Models omitting X1X_1DirectDR-NEAR
None0.52 ± 0.050.47 ± 0.07
Only the model used for graph construction0.52 ± 0.050.70 ± 0.07
First-stage outcome regressions and propensity network1.12 ± 0.051.04 ± 0.07

When the modifier is omitted from all first-stage models, both methods have higher error, but DR-NEAR still has lower mean error than Direct.

3. The results of shuffling and mixing show that weaker perturbations preserve DR-NEAR's advantage and stronger perturbations raise its mean error above Direct's.

We randomly shuffle the training plug-in CATE estimates that lie between their 25th and 75th percentiles, then replace the estimates used for graph construction within this interval by

\qquad s\in\{0.25,0.5,1\}.$$ Here, $p_i$ is the original estimate, $\pi$ is the random permutation within this interval, and $s$ is the mixing coefficient. Values outside the interval remain unchanged. Within each repetition, all three values of $s$ use the same permutation. We reconstruct the graph while keeping second-stage inputs and training settings fixed. Table D shows that DR-NEAR's mean error remains below that of Direct at $s=0.25$, but exceeds it at $s=0.5$ and $s=1$. Direct reuses its original predictions and remains at 0.52 ± 0.05. `Table D.` Shuffling and mixing plug-in CATE estimates in additional Synthetic experiments. | Estimates used for graph construction | DR-NEAR | | --- | ---: | | Original plug-in CATE estimates | 0.47 ± 0.07 | | $s=0.25$ | 0.50 ± 0.08 | | $s=0.5$ | 0.58 ± 0.06 | | $s=1$ | 0.61 ± 0.06 | **Thus, NEAR can improve on Direct by using relative effect information, but the additional experiments show that this benefit can disappear when graph construction omits an effect modifier or uses strongly perturbed effect orderings.** We again thank you for these valuable suggestions! In the revised manuscript, we have clarified the distinction from CIC in Section 2 and added the empirical comparison in Table 2. We have also clarified the connection between graph construction and the risk bound in Section 5.2 and the limitation due to first-stage estimation errors in Section 7. Your suggestions have really helped us present NEAR's contribution, its benefits, and the conditions under which these benefits disappear more clearly. We sincerely hope these responses address your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.