acceptodds
Under review as a conference paper at ICLR 2027

Joint Regression Adjustment for Sequential Randomized Experiments with Pretrained Predictions

Abstract

In longitudinal randomized experiments on internet platforms, researchers often need to assess the effects of sequential interventions on long-term outcomes. In such experiments, there are two types of potentially useful information for predicting outcomes: within-experiment covariates that evolve with past treatments, and fixed pretrained predictors trained using offline data available before the experiment. However, the former are affected by past treatments, while the latter may not transfer accurately to the current experiment. These difficulties limit the usability of such information. We propose a joint regression adjustment method for estimating average potential outcomes under fixed treatment paths and causal contrasts between these paths. It learns sequential outcome predictions through cross-fitting using time varying covariates within the experiment, then jointly adjusts for these predictions and fixed pretrained predictions. Under general conditions, we establish asymptotic normality and consistent variance estimation for valid Wald inference. Its asymptotic variance is no larger than that of Horvitz–Thompson, Hájek, and AIPW estimators. We characterize the incremental variance reduction from the pretrained predictions and give a necessary and sufficient condition for attaining the semiparametric efficiency bound, which can hold even when the internal predictions are misspecified. We demonstrate the advantages of our method through synthetic experiments and real-data experiments on a Chinese game platform.

Primary area:
causal reasoning
Keywords:
online experiments, longitudinal randomized experiments, covariate adjustment, causal inference, variance reduction, pretrained predictions, semiparametric inference

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

The paper focuses on estimating the potential outcome (and contrasts) at T, due to a sequence of binary treatment decisions from t=1, ..., T in a setting of sequential randomized experiments. In particular, they analyze an environment where pretrained predictions, mapping history at t=1, ..., T-1 to the final outcome, are available outside of the experiment. Using a Double ML approach, they develop an estimator that is robust against misspecification of the two outcome models (within experiment and before the experiment) and asymptotically normal. Furthermore, the asymptotic variance of the estimator approaches the semiparametric efficiency bound under some conditions and is no worse than baseline estimators that don't use pre-trained predictions such as Horvitz-Thompson and Hajek.

Critical strengths

Develops a method that incorporates pre-trained predictions for estimating the potential outcomes along some treatment path, where the source data comes from sequential experiments. There is a wide literature on dynamic treatment regimes and separately how to efficiently incorporate data outside on an experiment (say, pretrained predictors) to reduce variance for inference on data from an experiment. This extends the latter into the former. Crucially, the proposed estimator is asymptotically normal and first order robust to nuisances (i.e both outcome models). This latter point supports most of my score, reproducing the use of pre-trained predictors into the sequential treatment setting, while preserving asymptotic normality and robustness to misspecification in the outcome models. The variance decomposition is particularly clean, highlighting the gap between the semiparametric efficiency bound as well improvements from including the pre-trained predictors.

Furthermore, the experimental evidence is strong: lowest MSE across estimands, with near-nominal coverage; gains over AIPW/LTMLE using ff as a feature are consistent but modest. The only complaint is these are asymptotic results so some notion of finite sample guarantees would be nice.

Critical weaknesses

Finite-sample behavior is only partly examined. The T ablation stops at 7, and MSE for all methods more than doubles from T = 6 to T = 7 (1.33 -> 3.78 for joint RA), consistent with the on-path fraction shrinking exponentially. At n = 1000, standard-SE coverage drops to 86–88% and HC2 is needed (Table 11). With 1 + 3T regressors and Q^t,ft\hat{Q}_t, f_t likely near-collinear when effects are small, I’d like to see effective sample size or weight variance, condition numbers of the design, and MSE (not just coverage) at smaller n.

No real weaknesses otherwise, there was some minor missing definitions, e.g IkcI_k^c, that I had to puzzle out, also μ^o\hat{\mu}^o is defined and not used in the main text. In general I think the authors should use the 10 page limit to expand exposition in certain areas (e.g assumptions and potential violations etc).

The assumption re convergence in probability to a fixed QoQ^o mostly makes sense, especially for something like a linear regression, but its not obvious to me what a flexible learner like a GBDT would do as n→∞n \rightarrow \infty.

Minor:

  • Define ϵ^i\hat{\epsilon}_i at line 235 and IkcI_k^c at its first use.

The key reason for keeping the score to weak accept is novelty. It takes an existing axis of improvement (pretrained predictions on a larger historical dataset) already well developed in one domain (static experiments) to an extension of that domain, sequential experiments where assumptions are standard (e.g no major assumption is weakened)

Questions for authors

None. My assessment does not hinge on clarifications; suggestions are noted under weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer cHnX

Thank you for your careful reading and constructive feedback, and for recognizing the strengths of our theoretical analysis and empirical evaluation. We have revised the manuscript accordingly. All changes are highlighted in blue. Below, we respond to your comments.

1. Finite-sample performance and stability.

Thank you for the opportunity to further clarify the finite-sample performance of our method. We fully agree that observations following a specified treatment path become sparse as the number of stages increases, which is an important challenge in longitudinal causal inference. We have added effective sample sizes and design matrix condition numbers for experiments varying the sample size (Appendix C.6, Table 18) and the number of stages (Appendix C.7, Table 20). We also report MSE at smaller sample sizes (Appendix C.6, Table 17).

The larger condition numbers for path contrasts in the Lorenz experiment mainly arise because the two paths’ first-stage internal predictions are similar. Using the same replications, we further examined retaining only one of these prediction terms in the main setting, at a smaller sample size, and with a longer treatment sequence. The condition numbers decreased substantially, while relative changes in MSE were below 0.4% in all cases, and coverage and mean confidence interval length remained largely unchanged. These results support the stability of our method in the settings examined.

2. Notation and exposition.

Thank you for your careful reading. We have defined IkcI_k^c and ϵ^i\widehat{\epsilon}_i at their first use and moved μ^∘\widehat{\mu}^{\circ}, which was not used in the main text, to the proof.

3. Prediction convergence assumption.

Thank you for raising this question. Our assumption requires that, as the sample size increases, prediction variability induced by the randomness of the training data diminishes, and the internal predictions converge in L2L_2 to deterministic limits. These limits need not equal the true sequential outcome regressions, allowing persistent bias from regularization, model misspecification, or error propagation through recursive regressions. For flexible learners such as GBDT, verifying this condition does require further analysis of their complexity and training procedures. Compared with double machine learning inference that typically relies on product-rate conditions for estimation errors, we allow these misspecified limits and do not require a specified prediction convergence rate.

AuthorsOpenReview ↗

Response to Reviewer cHnX (continued)

4. Additional theoretical and methodological contributions.

We appreciate your recognition of the value of our work. Beyond extending results from single-stage to sequential experiments, we would like to explain three additional theoretical and methodological contributions:

(i) Inference with cross fitting under weaker prediction convergence conditions. The works most closely related to ours are Guo et al. (2021) and De Bartolomeis et al. (2025). Guo et al. (2021) require L4L_4 convergence and higher moment conditions, whereas we require only L2L_2 convergence of internal predictions to deterministic limits to establish asymptotic normality. The formal proof of De Bartolomeis et al. (2025) relies on independence between the training and estimation samples, and their Theorem 2 additionally assumes consistent covariance matrix estimation. We directly prove asymptotic normality and consistency of variance estimation for the cross-fitted estimator under our regularity conditions, without additionally assuming independence between the training and estimation samples or consistency of the covariance matrix estimator.

(ii) Precise efficiency characterization beyond a no-harm guarantee. De Bartolomeis et al. (2025) establish a no-harm variance guarantee for pretrained predictions. We further derive an exact variance decomposition by stage, quantify the incremental value of external predictions beyond internal adjustment, and give necessary and sufficient conditions for attaining the semiparametric efficiency bound even when internal predictions are misspecified.

(iii) Direct adjustment of sequential path contrasts. We directly optimize the asymptotic variance of the path contrast and remove exact collinearity induced by the design according to the first divergence stage, preserving the full adjustment space while establishing valid inference and variance guarantees.

Thank you again for your valuable feedback. We hope that our responses and the revisions address all of your questions and concerns. We would be happy to provide any further clarifications or address additional questions or suggestions.

References

Yongyi Guo, Dominic Coey, Mikael Konutgan, Wenting Li, Chris Schoener, and Matt Goldman. Machine learning for variance reduction in online experiments. In Advances in Neural Information Processing Systems, volume 34, 2021.

Piersilvio De Bartolomeis, Javier Abad, Guanbo Wang, Konstantin Donhauser, Raymond M. Duch, Fanny Yang, and Issa J. Dahabreh. Efficient randomized experiments using foundation models. In Advances in Neural Information Processing Systems, volume 38, pp. 159470–159509, 2025.

Appreciate the response and clarification on what were already minor points. No more questions from me!

AuthorsOpenReview ↗

Thank You for Your Feedback

Thank you very much for your thoughtful feedback. We sincerely appreciate your careful review and constructive suggestions, which have helped us strengthen the paper.

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

This paper proposes joint regression adjustment for estimating treatment-path means and contrasts in sequential randomized experiments. It combines cross-fitted internal predictions with fixed pretrained predictions to improve precision. The authors establish asymptotic normality, consistent variance estimation, and variance guarantees under prediction misspecification, and characterize conditions for semiparametric efficiency. Synthetic experiments and a gaming-platform A/A study, evaluated over repeated randomizations, demonstrate precision gains and approximately nominal confidence interval coverage.

Critical strengths
  1. The theoretical treatment is comprehensive, covering asymptotic normality, consistent variance estimation, variance comparisons, and conditions for semiparametric efficiency.

  2. The method addresses a practical problem: incorporating potentially misspecified pretrained predictions into sequential experiments. The final adjustment is a simple regression, making the procedure relatively easy to implement. The experiments use repeated evaluations 400 synthetic replications and 1,000 A/A replications—and report aggregate precision and coverage. This reduces reliance on individual runs and provides more reliable empirical evidence.

Critical weaknesses
  1. Limited experimental breadth is my main concern. The real-data evaluation uses a single gaming-platform A/A study, which does not test actual treatment-induced changes in subsequent behavior. An additional independent real-world dataset with genuine interventions would strengthen the evidence.The experiments omit the closely related REG/EMP methods of Kallus and Uehara (2019). A direct comparison would better establish the practical contribution beyond existing variance-optimized estimators.

  2. The benefit of jointly estimating a path contrast is not directly isolated. An ablation comparing joint adjustment with separately adjusted path means would help substantiate this component.The distinction from existing regression-adjustment methods could be sharper. Although the inference theory is well developed, the paper should more clearly separate its new contributions from established control-variate and variance-minimization principles.

Questions for authors
  1. Is there evidence that the gains extend beyond the single-platform A/A setting to experiments with actual interventions?

  2. How does the method compare with REG/EMP, and how much improvement comes specifically from jointly adjusting the path contrast?

  3. In Appendix B, does the A/A simplification explicitly enforce shared predictions across paths? Equality of the true regressions alone need not imply equality of fitted predictions.

4 comments
AuthorsOpenReview ↗

Response to Reviewer FpLV

Thank you for your careful reading and constructive feedback, and for recognizing the strengths of our theoretical analysis and empirical evaluation. We have revised the manuscript accordingly. All changes are highlighted in blue. Below, we respond to each of your questions.

Question 1 Response

Thank you for this suggestion. The true treatment effect in the A/A experiment is zero, allowing us to assess MSE and coverage under a null scenario. We have added a semisynthetic sequential experiment based on real customer covariates from Hillstrom (2008). This allows us to evaluate our method under nonzero treatment effects while preserving the distribution of real customer characteristics.

Our method achieves the lowest empirical MSE for both τ1\tau^1 and τ2\tau^2. The 95% confidence intervals attain empirical coverage of 94.75% for τ1\tau^1 and 95.75% for τ2\tau^2 (Section 6, “Semisynthetic experiment,” Table 2). The full data-generating mechanism and configurations are provided in Appendix D.

Question 2 Response

Thank you for suggesting these comparisons and the ablation to assess the additional gains from joint adjustment.

1. Comparison with REG and EMP. We have added comparisons with REG and EMP (Kallus and Uehara, 2019) in all three experiments; full results are provided in Appendix F. Our method reduces empirical MSE relative to REG and EMP, respectively, by:

  • Lorenz: 8.4% and 8.8% for τ1\tau_1; 8.6% and 8.7% for τ2\tau_2. (Appendix F, Table 30).
  • Hillstrom: 34.2% and 34.7% for τ1\tau^1; 5.2% and 7.1% for τ2\tau^2. (Appendix F, Table 31).
  • Real A/A: 7.1% and 7.7% with Transformer; 5.1% and 5.5% with LightGBM. (Appendix F, Table 32).

2. Additional gains from joint adjustment. Let εaˉ\varepsilon^{\bar a} and εbˉ\varepsilon^{\bar b} denote the population residuals from separate path adjustment, and let CτC^\tau contain the joint adjustment variables after removing redundancies. Define g=E[Cτ(εaˉ−εbˉ)]g=E[C^\tau(\varepsilon^{\bar a}-\varepsilon^{\bar b})] and Σ=E[Cτ(Cτ)⊤]\Sigma=E[C^\tau(C^\tau)^\top]. Under the corresponding regularity conditions, the asymptotic variance reduction is Vsep,τ−Vτ=g⊤Σ−1g≥0V_{\mathrm{sep},\tau}-V_\tau=g^\top\Sigma^{-1}g\geq0. Thus, direct adjustment can further reduce variance when a residual from one path remains predictable from the other path's adjustment variables.

In the new Hillstrom semisynthetic experiment, we compare direct adjustment of the path contrast with separately adjusting the two path means and subtracting, for two path-contrast estimands with shared prefixes. This comparison evaluates the additional gains from joint adjustment.

Table 1: Hillstrom joint and separate adjustment over the same 400 seeds. MSE is multiplied by 10210^2 and mean CI length by 1010. Parentheses contain MCSEs. Coverage is reported as a percentage.

PredictionsAdjustmentτ1\tau^1 MSEτ1\tau^1 Cov.τ1\tau^1 Lengthτ2\tau^2 MSEτ2\tau^2 Cov.τ2\tau^2 Length
QQSeparate38.11 (2.62)95.0025.45 (0.11)17.92 (1.24)94.7517.34 (0.08)
Joint31.11 (2.08)94.0022.86 (0.09)17.69 (1.25)95.2517.36 (0.09)
Q+fQ+fSeparate36.55 (2.68)94.2524.44 (0.12)16.28 (1.15)95.5016.41 (0.06)
Joint22.47 (1.70)94.7518.44 (0.08)14.47 (1.09)95.7515.65 (0.05)

As shown in Table 1, direct joint adjustment reduces MSE by 38.5% and 11.2% with external predictions, and by 18.4% and 1.3% using internal predictions alone, relative to separate adjustment.

References

Kevin Hillstrom. The MineThatData e-mail analytics and data mining challenge, 2008.

Nathan Kallus and Masatoshi Uehara. Intrinsically efficient, stable, and bounded off-policy evaluation for reinforcement learning. In Advances in Neural Information Processing Systems, volume 32, 2019.

AuthorsOpenReview ↗

Response to Reviewer FpLV (continued)

Question 3 Response

We fully agree that equality of population regressions does not imply equality of fitted predictions in finite samples. The original A/A simplification in Appendix B was intended to avoid duplicate prediction terms. Our implementation retains the internal prediction adjustment terms for both paths, and the specific adjustment terms are now explicitly listed in Appendix E.1. For comparison, we also evaluated retaining both internal prediction adjustment terms versus retaining only one at the first stage, using the same 1,000 replications. As shown below, the numerical differences between the two implementations are small.

Table 2: Sensitivity to the first-stage internal prediction terms in the real A/A experiment. Parentheses contain MCSEs. Coverage is reported as a percentage.

External predictorFirst-stage termsEmpirical varianceMSECov.Mean CI length
TransformerBoth (original)69.5469.49 (3.02)95.3032.73 (0.11)
One69.5169.46 (3.02)95.2032.73 (0.11)
LightGBMBoth (original)70.9470.89 (3.02)95.0032.67 (0.10)
One70.8970.84 (3.01)94.8032.67 (0.10)

Thank you again for your valuable feedback. We hope that our responses and the revisions address all of your questions and concerns. We would be happy to provide any further clarifications or address additional questions or suggestions.

Thanks for your detailed response! Your answers have addressed most of my concerns! I will maintain my postive score and improve my confience to 4!

AuthorsOpenReview ↗

Thank You for Your Feedback and Support

Thank you very much for your positive follow-up and for taking the time to carefully review our revisions. We are glad that the additional experiments and clarifications have addressed most of your concerns. Your constructive suggestions have helped us substantially strengthen the paper. We sincerely appreciate your time and feedback!

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

This paper studies estimation and inference of treatment-path means and causal contrasts in sequential randomized experiments using both internal outcome predictions and pretrained predictions. They proposed a joint regression adjustment framework that consists of three steps: 1. Sequential outcome prediction, 2. Joint regression adjustment, and 3. Confidence intervals construction. The authors establish asymptotic normality, consistent variance estimation, and asymptotic variance guarantees for the proposed framework. They also compared it with other existing estimators through both simulation studies and a real-data experiment using gaming platform data.

Critical strengths
  1. The theoretical analysis is solid. The authors propose a joint regression adjustment framework that combines internal outcome predictions with external pretrained predictions for efficient estimation and inference of treatment-path means and causal contrasts in sequential randomized experiments. They establish asymptotic normality, consistency of the variance estimator, and guarantees on asymptotic variance.

  2. The empirical evaluation includes both simulations and a real-data application. The results demonstrate that the proposed framework can achieve efficiency gains over CF-AIPW and CF-LTMLE in the settings considered.

Critical weaknesses
  1. The methodological novelty is limited. The authors extend existing literature that use external pretrained predictors to improve efficiency in randomized experiments from single-stage to longitudinal settings. While this extension is useful, the contribution appears incremental rather than a substantial methodological advance.

  2. The empirical gains are limited. In the noise-sensitivity analysis, the advantage over CF-AIPW and CF-LTMLE disappears even at a noise level of $ \lambda=0.05 $, raising concerns about the robustness of the reported gains.

  3. The authors report MSE and average confidence interval length without quantifying the Monte Carlo uncertainty in these metrics. It is therefore unclear whether the observed improvements reflect meaningful gains or simulation variability.

  4. The authors do not provide sufficient details about the pretrained transformer-based regression predictor used in the real-data analysis.

Questions for authors
  1. Can the author quantify the uncertainty when reporting MSE and average confidence interval length?

  2. Can the author provide sufficient details about the pretrained transformer-based regression predictor used in the real-data analysis and provide some sensitivity analysis of the choice of the pretrained predictors?

3 comments
AuthorsOpenReview ↗

Response to Reviewer onLr

Thank you for your careful reading and constructive feedback. We have revised the manuscript accordingly. All changes are highlighted in blue. Below, we respond to each weakness.

Weakness 1 Response

We thank the reviewer for the opportunity to clarify our contributions beyond existing work on single-stage experiments. In sequential experiments, earlier treatments can affect subsequent covariates. As Rosenbaum (1984) pointed out, directly adjusting for treatment-affected variables as ordinary covariates can introduce bias. To solve the problem, we combine each stage-specific prediction with the corresponding path-weight increment, using E[(W2aˉ−W1aˉ)f2(H2)∣H2]=0E[(W_2^{\bar a}-W_1^{\bar a})f_2(H_2)\mid H_2]=0, guaranteed by sequential randomization, to construct mean-zero adjustment terms. This allows us to use time-varying information affected by earlier treatments to improve precision while preserving the original treatment-path estimand. Building on this construction, we make the following additional theoretical and methodological contributions beyond existing single-stage work:

  1. Inference with cross fitting under weaker prediction convergence conditions. The works most closely related to ours are Guo et al. (2021) and De Bartolomeis et al. (2025). Guo et al. (2021) require L4L_4 convergence and higher moment conditions, whereas we require only L2L_2 convergence of internal predictions to deterministic limits to establish asymptotic normality. The formal proof of De Bartolomeis et al. (2025) relies on independence between the training and estimation samples, and their Theorem 2 additionally assumes consistent covariance matrix estimation. We directly prove asymptotic normality and consistency of variance estimation for the cross-fitted estimator under our regularity conditions, without additionally assuming independence between the training and estimation samples or consistency of the covariance matrix estimator.

  2. Precise efficiency characterization beyond a no-harm guarantee. De Bartolomeis et al. (2025) establish a no-harm variance guarantee for pretrained predictions. We further derive an exact variance decomposition by stage, quantify the incremental value of external predictions beyond internal adjustment, and give necessary and sufficient conditions for attaining the semiparametric efficiency bound even when internal predictions are misspecified.

  3. Direct adjustment of sequential path contrasts. We directly optimize the asymptotic variance of the path contrast and remove exact collinearity induced by the design according to the first divergence stage, preserving the full adjustment space while establishing valid inference and variance guarantees.

References

Paul R. Rosenbaum. The consequences of adjustment for a concomitant variable that has been affected by the treatment. Journal of the Royal Statistical Society: Series A (General), 147(5):656–666, 1984.

Yongyi Guo, Dominic Coey, Mikael Konutgan, Wenting Li, Chris Schoener, and Matt Goldman. Machine learning for variance reduction in online experiments. In Advances in Neural Information Processing Systems, volume 34, 2021.

Piersilvio De Bartolomeis, Javier Abad, Guanbo Wang, Konstantin Donhauser, Raymond M. Duch, Fanny Yang, and Issa J. Dahabreh. Efficient randomized experiments using foundation models. In Advances in Neural Information Processing Systems, volume 38, pp. 159470–159509, 2025.

AuthorsOpenReview ↗

Response to Reviewer onLr (continued)

Weakness 2 Response

Thank you for the opportunity to further explain the empirical advantages of our method. To compare our method with the other methods more clearly, we report differences in MSE and mean confidence interval length, together with their Monte Carlo standard errors computed across the same simulation replications.

At λ=0.05\lambda=0.05, for the path means, our method and the two baselines have similar MSEs; for the path contrasts τ1\tau_1 and τ2\tau_2, our method reduces MSE by 5.2%–5.5% and 1.9%–2.5%, respectively. Across the noise levels and estimands considered, our method has lower empirical MSE in 65/72 comparisons. In this setting, our method also offers an advantage in confidence interval length: mean confidence intervals are shorter in all 72 comparisons, with statistically significant reductions in 64/72, including all comparisons for both path contrasts (Appendix C.5, Table 12).

Besides noisy external predictions, training pretrained predictors on external data under distribution shift and using a different internal learner are two other important settings. Our method is broadly competitive in both:

1. Gains under distribution shift. The different γ,κ\gamma,\kappa settings simulate a shift between the external predictor's training distribution and the experimental distribution, which is common in practice. Our method has lower empirical MSE in 144/144 comparisons, with 129/144 reductions statistically significant. Its mean confidence intervals are significantly shorter in all comparisons (Appendix C.5, Tables 13–15).

2. Gains with a different internal learner. Using MLP for internal sequential outcome prediction in the main setting, our method reduces MSE by 30.3%–43.3% relative to the two baselines. Both the MSE reductions and the shorter mean confidence intervals are statistically significant for all four estimands (Appendix C.5, Table 16).

Weakness 3 Response (Question 1)

We fully agree that quantifying uncertainty in the reported results is important. We have added Monte Carlo standard errors for MSE and mean confidence interval length to all tables throughout the revised manuscript.

Weakness 4 Response (Question 2)

Thank you for this comment. We have added details of the external predictors' training data, training procedures, and model configurations in Appendix E.1, including the Transformer's number of layers, hidden dimension, and attention heads. We also include a LightGBM external predictor and report sensitivity results for both predictors across different noise levels in Appendix E.2.

Using either external predictor, our method maintains coverage close to 95%. Differences between methods, together with their Monte Carlo standard errors, show that our method significantly reduces empirical variance and mean confidence interval length relative to CF-AIPW and CF-LTMLE using the same external predictions:

Comparisons using Transformer and LightGBM external predictors over 1,000 replications. Reductions are computed as the comparator minus our method. Monte Carlo standard errors are shown in parentheses.

External predictorComparisonVariance reductionMean CI length reduction
TransformerCF-AIPW vs. ours4.52 (1.41)0.88 (0.02)
CF-LTMLE vs. ours4.68 (1.30)0.71 (0.02)
LightGBMCF-AIPW vs. ours3.92 (1.46)0.95 (0.02)
CF-LTMLE vs. ours3.73 (1.34)0.79 (0.02)

Thank you again for your valuable feedback. We hope that our responses and the revisions address all of your questions and concerns. We would be happy to provide any further clarifications or address additional questions or suggestions.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer onLr,

Thank you again for your review. Could you let us know whether our revision and rebuttal have addressed your concerns? We are happy to clarify any remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.