- 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 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 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 , that I had to puzzle out, also 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 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 .
Minor:
- Define at line 235 and 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
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 convergence and higher moment conditions, whereas we require only 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!
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.
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 Ikc and ϵi at their first use and moved μ∘, 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 L2 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.