Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read OPTrack, an optimistic adaptive sampling rule, achieves logarithmic Neyman regret against the minimum variance attainable by any estimator and allocation, whereas prior nonasymptotic algorithms only competed with IPW baselines.

desk verdict Genuinely new optimistic allocation for adaptive ATE estimation with a plausible stronger-baseline log regret, but the proof has a real per-arm variance CS gap and a few typos that need fixing. read the letter →

arxiv 2502.04673 v1 pith:I2YGI4N2 submitted 2025-02-07 stat.ML cs.LGstat.ME

Ojash Neopane , Aaditya Ramdas , Aarti Singh This is my paper · ORCID
classification stat.MLcs.LGstat.ME
keywords adaptiveexperimentaldesignaveragetreatmenteffectNeymanregretallocationAIPWestimatoroptimismprincipleconfidencesequencessequentialestimation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes OPTrack, an adaptive sampling algorithm for estimating the average treatment effect (ATE) when treatment assignments are chosen online. Its central claim is that OPTrack's Neyman regret—the normalized squared error compared with the lowest variance attainable by any estimator–allocation pair—grows only logarithmically in the horizon $T$, with a rate depending on the gap between the two arms' standard deviations. Earlier nonasymptotic algorithms measured regret against weaker IPW baselines and incur linear regret under this stronger definition; OPTrack is the first finite-sample analysis to target the AIPW baseline directly. The authors show empirically that the improvement translates into roughly 10–15 percent lower normalized MSE in small samples and competitiveness with an oracle that knows the Neyman allocation but must still learn the rewards.

What carries the argument

The central object is the confidence sequence $\mathrm{CS}_t(\pi^\star)$ for the Neyman allocation, built from per-arm confidence sequences for the standard deviations $\sigma(a)$. Lemma B.1 supplies a time-uniform interval around the empirical standard deviation with width of order $\sqrt{(\log\log t + \log(1/\delta))/t}$, and equation (10) feeds the lower and upper endpoints into the Neyman fraction to obtain $\mathrm{CS}_t(\pi^\star)$. OPTrack then sets $\pi_t = \arg\min_{\pi \in \mathrm{CS}_t(\pi^\star)} |1/2 - \pi|$. The optimism principle is what does the work: it forces the played allocation to stay on the Neyman side of $1/2$ once exploration ends, so the per-round Neyman loss is controlled by $|\pi_t - \pi^\star|^2$ plus the reward-estimation error, both of which shrink like $1/(\pi^\star t)$.

What would settle it

Fix a two-arm Bernoulli environment with known $\sigma(0) \neq \sigma(1)$, run OPTrack for horizons $10^3, 10^4, 10^5$, and plot the empirical normalized Neyman regret $T \cdot \mathrm{MSE} - V^\star$. If regret grows linearly in $T$, or its constant grows faster than $\log T$, the Theorem 5.3 bound fails. A sharper check is to test coverage of the variance intervals at stopping times where one arm's count $N_t(a)$ is small; coverage below $1-\delta$ would locate the failure in the transfer from full-sample to random-denominator confidence sequences.

Watch

Extended reading notes

Core claim

The paper establishes that optimism can resolve the exploration–exploitation tradeoff in adaptive ATE estimation. The proposed OPTrack algorithm maintains a confidence sequence for the unknown Neyman allocation $\pi^\star = \sigma(1)/(\sigma(0)+\sigma(1))$, and each round plays the allocation in that confidence set that is closest to $1/2$. Because the problem becomes harder as the Neyman allocation moves away from $1/2$, playing the point closest to $1/2$ is acting as if the environment were the easiest feasible one—an instance of the optimism principle. Once $1/2$ leaves the confidence set, optimism keeps the played allocation on the Neyman side of $1/2$, which balances the two arms' sampling counts and yields, with probability at least $1-\delta$, the bound $\widetilde{O}(\Delta(\sigma)^{-2} + (1/\pi^\star)^2 \log T)$ on Neyman regret. The decisive contrast with past work is the baseline: regret is measured against the minimum variance over all estimator–allocation pairs, so the IPW-based ClipSMT and ClipOGD algorithms exhibit linear regret under the same measure.

Load-bearing premise

The regret guarantee rests on the per-arm variance confidence intervals remaining valid simultaneously over all rounds even though the denominator in each arm's sample count is random and data-dependent; the proof supplies a fixed-sample-size lemma and applies it to those random counts.

Editorial extensions

If this is right

  • OPTrack attains logarithmic Neyman regret against the minimum-variance efficiency baseline, so its mean squared error converges to the best attainable variance at rate $1/T$ rather than stagnating at a constant gap.
  • Because the regret bound depends on $1/\pi^\star$ squared, the algorithm adapts to problem difficulty: a Neyman allocation far from $1/2$ produces a short exploration phase and a small constant, while nearly balanced arms incur a longer exploration phase.
  • If Theorem 5.3 is correct, OPTrack strictly dominates the IPW-based algorithms ClipSMT and ClipOGD, which suffer linear regret under the stronger estimator–allocation baseline.
  • The A2IPW estimator used by OPTrack is compatible with anytime-valid confidence sequences, so the same allocation rule can feed sequential testing and adaptive inference without restarting exploration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The optimism principle suggests a template for adaptive estimation beyond ATE: whenever an optimal design is a functional of unknown nuisance parameters, playing the design closest to the 'easiest' feasible instance may convert design regret into logarithmic regret; the paper hints at this for multi-arm and covariate settings but does not prove it.
  • A natural testable extension is to probe tightness of the $1/\pi^\star$ squared dependence by varying the ratio $\sigma(0)/\sigma(1)$ in Bernoulli environments and measuring whether the empirical regret constant scales with the predicted factor.
  • The random-denominator transfer in the variance confidence sequence could be checked directly: if coverage of the Lemma B.1 intervals fails at stopping times where one arm's count $N_t(a)$ is small, a fully time-uniform per-arm confidence sequence built on stopped martingales would be the safer machinery.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes OPTrack, a fully adaptive algorithm for estimating the average treatment effect with the A2IPW estimator, and analyzes its Neyman regret in a nonasymptotic superpopulation setting (Section 3). The algorithm builds confidence sequences for the per-arm standard deviations, converts them into a confidence sequence CS_t(π*) for the Neyman allocation (Eq. 10), and plays the allocation in CS_t(π*) closest to 1/2 (Algorithm 1), interpreting this as optimism in the face of uncertainty. The main claim (Theorem 5.3) is that with probability at least 1−δ, OPTrack's Neyman regret is Õ(Δ(σ)^{-2} + (1/π*)^2 log T), where the baseline is the minimum variance over all estimator–allocation pairs, a strictly stronger benchmark than the IPW-optimal baselines used by prior nonasymptotic work (Dai et al., 2023; Neopane et al., 2025). The analysis splits time into an exploration phase whose length is bounded in Lemma 5.1 by Õ(Δ(σ)^{-2} log(1/δ)) and a concentration phase in which Lemma 5.2 bounds |π_t − π*| by Õ(√(ℓ(t,δ)/(π* t))). Experiments on six Bernoulli instances over 500,000 simulations report a 10–15% improvement in normalized MSE over ClipSDT at small T and competitiveness with an oracle that knows π*.

Significance. The manuscript targets a real gap: prior finite-sample analyses of adaptive ATE estimation compete against the minimum-variance IPW baseline, whereas the minimum variance over all estimator–allocation pairs is the appropriate semiparametric benchmark. The optimism principle — play the feasible allocation nearest 1/2 — is simple, interpretable, and eliminates the clipping-sequence hyperparameter that burdens ClipSDT, ClipSMT, and ClipOGD, giving a genuinely parameter-free procedure; this is a substantive algorithmic contribution. The experiments are transparent, use a defensible oracle baseline, and show consistent finite-sample gains. However, the main result is not yet verified as written: the regret defined in Eq. (5) is not the quantity bounded in Appendix A.1 (the two differ by a factor of T), the per-arm variance confidence sequence underpinning the good event (16) is not implied by the stated Lemma B.1 for random sample sizes, and several load-bearing bounds are imported without proof from the authors' companion paper. These are repairable correctness gaps rather than refutations of the approach, but they must be closed before the theoretical claims can be accepted.

major comments (4)
  1. [Section 3, Eqs. (5)–(6), and Appendix A.1] The object bounded in the proof is not the object defined as Neyman regret. With Var(bΔ_T) = (1/T^2) Σ_{t=1}^T E[ℓ(π_t, br_t)] (Lemma C.1), Eq. (5)'s normalized gap is (1/T) Σ_t E[ℓ_t] − V*, where V* = σ²(1)/π* + σ²(0)/(1−π*), while Eq. (6) defines R_T = Σ_t E[ℓ_t] − V*. The proof of Theorem 5.3, Eqs. (19)–(24), bounds Σ_t (E[ℓ_t] − V*), subtracting V* inside each round, so the exploration phase alone contributes 4T and the concentration phase contributes Σ 625ℓ(t,δ)/((π*)² t); these match the stated rate Õ(Δ(σ)^{-2} + (1/π*)² log T) only for the cumulative object Σ_t(E[ℓ_t] − V*). Thus the theorem's bound does not apply to the definition in Eq. (5) (where the same proof would imply only Õ((Δ(σ)^{-2} + (1/π*)² log T)/T)), and Eq. (6) is off by an additive (T−1)V* from the proof's quantity. The definition of R_T and the proof must be reconciled before the logarithmic-regret claim can be evaluated.
  2. [Appendix B (Lemma B.1) and Eq. (16)] The good event E_σ(δ̃) is asserted as |bσ_t(a) − σ(a)| ≤ 4.2√(ℓ(t,δ)/t) for all arms and all t, but Lemma B.1 provides a confidence sequence only for a variance estimate computed from all t observations of a single [0,1]-bounded sequence. In OPTrack, bσ_t(a) is computed from N_t(a) observations, where N_t(a) is a random sample size, so the natural conclusion of Lemma B.1 applied to the arm-a subsequence is |bσ_t(a) − σ(a)| ≤ 4.2√(ℓ(N_t(a),δ)/N_t(a)) for all t, and no argument converts that random-denominator event into the fixed-denominator bound (16); when N_t(a) is well below t, the displayed bound is stronger than what Lemma B.1 supplies. The crux is load-bearing because the proofs in Appendix A actually use the N_t(a) form — Z_t(a) in Lemma A.1 and Eq. (33) in Lemma 5.1 — while the union bound establishing E(δ̃) uses the t-denominator form. The repair of reindexing arm-specific observation times and stating a time-uniform per-arm CS is standard but is not written down. In addition, Lemma B.1 requires [0,1]-bounded rewards, an assumption absent from Section 3 and from the statement of Theorem 5.3, and the reward event E_r(δ) in Eq. (18) is displayed with √(tℓ(t,δ)) although the proof (Eq. (29)) uses √(ℓ(t,δ)/N_t(a)); both mismatches need to be corrected for the high-probability event to support the main theorem.
  3. [Section 5 and Appendix A] The proof of Theorem 5.3 imports four load-bearing results from the authors' companion paper Neopane et al. (2025) without stating them: Lemma 4.3 in Eq. (28), Lemma 4.5 in Eq. (31), Lemma B.10 in Eqs. (35)–(36) and in Lemma A.1, and Lemma 2 in the concentration-phase argument described after Theorem 5.3. These bounds carry the constant factors and the π*-dependence that determine the claimed rate, so the theorem cannot be checked from this manuscript alone. Even if the companion paper is correct, this dependence should be made explicit and the relevant lemmas stated (or proved) in an appendix; the current manuscript gives no way for a reader to verify that the cited results apply to the random-allocation, per-arm setting used here.
  4. [Appendix A.3, Eqs. (48)–(52)] The chain used to lower-bound the denominator in Eq. (47) is invalid as written. Line (48) states Z_{0,t} = 4.2√(ℓ(t,δ)/t), but with Z_{0,t} defined as 4.2√(ℓ(t,δ)/N_t(0)) and N_t(0) ≤ t, the quantity 4.2√(ℓ/t) is a lower bound, not an upper bound, on Z_{0,t}. Line (49), Z_{0,t} ≤ Δ(σ)/8, does not follow from t ≥ T: the exploration-phase exit condition (33) implies only Z_{0,t} + Z_{1,t} < Δ(σ) on the good event, with no split that gives either term a 1/8 margin. Line (50) then replaces Z_{1,t} in the denominator by −Δ(σ)/8 without justification. Since this chain supplies the inequality σ(0) + σ(1) − Z_{0,t} ≥ (σ(0)+σ(1))/2 used to obtain the constant 25 in Lemma 5.2 (Eq. (53)), a central ingredient of the concentration-phase bound of Theorem 5.3 rests on a proof step that does not hold as written; the lemma may be repairable through a different argument, but the current derivation needs to be reworked.
minor comments (6)
  1. [Section 1, Organization paragraph] The sentences 'The rest of our paper is organized as follows.' and 'The remainder of this paper is structured as follows.' are duplicated back-to-back, and the resulting list omits the Conclusion section that the paper actually contains.
  2. [Section 5, Notation] The notation 'we let π = min{π, 1 − π}' overloads the allocation π with its distance to the boundary; a distinct symbol (e.g., π̄) would avoid confusion in Lemma 5.1 and Theorem 5.3.
  3. [Eq. (18)] The displayed event E_r(δ), |br_t(a) − r*(a)| ≤ √(tℓ(t,δ)), is dimensionally inconsistent with its use in Eq. (29), which requires √(ℓ(t,δ)/N_t(a)); as printed the bound is also vacuous for bounded rewards.
  4. [Algorithm 1 and Eq. (10)] The lower endpoints L_t(σ(a)) from Lemma B.1 can be negative at small t (bσ_t − 1.7√(ℓ/t)); the algorithm should intersect CS_t(π*) with [0,1] or clip the endpoint bounds, otherwise the argmin in step 3 may range outside the unit interval. The timing of the update (CS_t uses data up to t−1) should also be stated explicitly.
  5. [Sections 2.1 and 6] The claim that ClipSMT and ClipOGD incur linear Neyman regret under the new baseline is asserted but never demonstrated; a one-line comparison of their limiting normalized MSE against V* would substantiate the exclusion from the experiments.
  6. [Eq. (7)] The term ε²_t(a) inside the definition of ℓ(π, r) carries a time index inappropriate for a generic loss function; the subscript should be dropped or made local to the definition.

Circularity Check

2 steps flagged · score 4.0 of 10

The central log T concentration-phase regret is imported from the authors' own companion paper via Lemma 2, and the proof also leans on companion lemmas 4.3, 4.5, and B.10; however, no fitted input is renamed as a prediction, so the result is not definitionally circular.

  1. self citation load bearing [Section 5, after Theorem 5.3; Appendix A.1, around equation (28)]
    "The second term in our bound is the Neyman regret during the concentration phase and follows from the application of Lemma 5.2 in conjunction with Lemma 2 of (Neopane et al., 2025) showing that the Neyman regret scales according to |π⋆ − πt|2 ≈ 1/π⋆·t."

    The main logarithmic term of Theorem 5.3 is not derived in this manuscript; it is supplied by Lemma 2 from a companion paper by the same authors. The present paper does not reproduce or verify that lemma, and it is the only support for converting the allocation-gap bound of Lemma 5.2 into a Neyman-regret bound. This makes the central claim load-bearing on a self-citation rather than on a derivation contained in the present manuscript. It is not an equality-by-construction and no fitted parameter is renamed, so it raises the score without making the whole result definitionally circular.

  2. self citation load bearing [Appendix A.2, proof of Lemma 5.1; Appendix A.3, proof of Lemma 5.2]
    "Applying Lemma B.10 from (Neopane et al., 2025) shows that whenever t ≥ eO(log(1/δ)), we have that 1.7 sqrt(ℓ(t,δ)/t) < 1/4 so that we need to bound min { t : t > 64/Δ^2(σ) · ℓ(t,δ) }."

    The exploration-time bound in Lemma 5.1 and the allocation-convergence rate in Lemma 5.2 both depend on Lemma B.10, and the per-round regret bound in Appendix A.1 also invokes Lemmas 4.3 and 4.5 from the same companion paper, all cited without proof. These self-citations are load-bearing because, without them, the displayed bounds on T and |π_t − π⋆| are not established by the present analysis. The cited statements are separate mathematical lemmas rather than definitions of the target result, so this is dependency through self-citation rather than a purely constructional circle.

full rationale

OPTrack is benchmarked against the external Neyman variance and uses the published A2IPW estimator, so the central comparison is not defined in terms of the algorithm's own outputs. No fitted parameter is renamed as a prediction, and no quantity is defined to be equal to the claimed result by construction. The main circularity concern is the heavy reliance on the authors' companion paper: Theorem 5.3's log T concentration term is attributed to Lemma 2 of Neopane et al. (2025), and the supporting Lemmas 5.1 and 5.2 use Lemma B.10, 4.3, and 4.5 from the same companion work without re-derivation. These are parameter-free mathematical lemmas with stated assumptions, so they are real evidence rather than vacuous self-citation, which prevents a higher score; still, the present paper is not self-contained for its central claim. Separately, there is a correctness gap rather than a circularity: the good event Eσ(δ) in equation (16) uses a confidence width with denominator t, while the algorithm's variance estimate bσ_t(a) has random denominator N_t(a), so the high-probability event as written is not formally justified; this gap is weighed in the verdict but is not an equivalence-by-construction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim is grounded in an external benchmark, the Neyman variance, and a published estimator, A2IPW. The proof outsources several technical bounds to the authors' own companion paper, but no free parameters or invented entities are introduced.

assumptions (4)
  • domain assumption Unconfoundedness and fixed conditional reward means and variances: E[Rt(a)|Ft-1]=r*(a) and V[Rt(a)|Ft-1]=σ^2(a) for all t.
    The A2IPW variance decomposition and the Neyman allocation in Section 3 rely on the potential outcome means and variances being constant across rounds.
  • domain assumption Rewards are bounded in [0,1].
    Used in Lemma B.1 for the variance confidence sequence and in bounding ε_t(a) ≤ 1 in the exploration phase proof.
  • standard math Time-uniform martingale concentration results of Howard et al. (2021), specifically Theorem 1 for the good event and Lemma B.1.
    The paper invokes these results to control the events Eσ, EN, and Er in Appendix A, treating them as unproved background.
  • ad hoc to paper Lemmas 4.3, 4.5, B.10 and Lemma 2 from Neopane et al. (2025) are correct and applicable.
    The proofs of Theorem 5.3 and Lemma 5.2 rely on these bounds, which are imported from the authors' own companion paper and not re-derived in this submission.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect." pith.science (2026). https://pith.science/paper/I2YGI4N2

@misc{pith2026250204673,
  author       = {Pith},
  title        = {Pith review of: Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I2YGI4N2}},
  note         = {Machine review of arXiv:2502.04673}
}
read the original abstract

Estimation and inference for the Average Treatment Effect (ATE) is a cornerstone of causal inference and often serves as the foundation for developing procedures for more complicated settings. Although traditionally analyzed in a batch setting, recent advances in martingale theory have paved the way for adaptive methods that can enhance the power of downstream inference. Despite these advances, progress in understanding and developing adaptive algorithms remains in its early stages. Existing work either focus on asymptotic analyses that overlook exploration-exploitation tradeoffs relevant in finite-sample regimes or rely on simpler but suboptimal estimators. In this work, we address these limitations by studying adaptive sampling procedures that take advantage of the asymptotically optimal Augmented Inverse Probability Weighting (AIPW) estimator. Our analysis uncovers challenges obscured by asymptotic approaches and introduces a novel algorithmic design principle reminiscent of optimism in multiarmed bandits. This principled approach enables our algorithm to achieve significant theoretical and empirical gains compared to prior methods. Our findings mark a step forward in advancing adaptive causal inference methods in theory and practice.

Figures

Figures reproduced from arXiv: 2502.04673 by the authors.

Figure 1
Figure 1. A plot of ℓ(π, r⋆ ) where π ⋆ < 1 2 . Note how the Neyman loss is smaller for π ⋆ + ϵ. This is due to the fact that π ⋆ + ϵ is closer to 1 2 , highlighting how less explorative allocations incur larger Neyman regret. mean rbt (a) = 1 Nt−1(a) Pt−1 s=1 Rs · I[At = a]. The main difficulty now is in constructing the confidence sequence CSt(π ⋆ ). In order to do so, we first construct confidence sequences for the standar… view at source ↗
Figure 2
Figure 2. Normalized MSE (T · MSE) for OPTrack, ClipSDT, and the oracle baselines across six problem instances, each with Bernoulli rewards with µ1 = 1 2 and varying µ0. Results are averaged over 500,000 simulations. OPTrack consistently outperforms ClipSDT, with a 10-15 improvement for smaller T. Notably, OPTrack is competitive with the reward estimation oracle and even outperforms it in some cases due to better exploration … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Use of variance estimation in the multi-armed bandit problem

    Audibert, J.-Y., Munos, R., and Szepesv \'a ri, C. Use of variance estimation in the multi-armed bandit problem. 2006

    work page 2006

  3. [3]

    Semiparametric efficient inference in adaptive experiments

    Cook, T., Mishler, A., and Ramdas, A. Semiparametric efficient inference in adaptive experiments. In Proceedings of the Third Conference on Causal Learning and Reasoning, volume 236 of Proceedings of Machine Learning Research, pp.\ 1033--1064. PMLR, 01--03 Apr 2024

    work page 2024

  4. [4]

    Clip- OGD : An experimental design for adaptive N eyman allocation in sequential experiments

    Dai, J., Gradu, P., and Harshaw, C. Clip- OGD : An experimental design for adaptive N eyman allocation in sequential experiments. Advances in Neural Information Processing Systems, 36: 0 32235--32269, 2023

    work page 2023

  5. [5]

    Minimax-optimal off-policy evaluation with linear function approximation

    Duan, Y., Jia, Z., and Wang, M. Minimax-optimal off-policy evaluation with linear function approximation. In International Conference on Machine Learning, pp.\ 2701--2709. PMLR, 2020

    work page 2020

  6. [6]

    Doubly robust policy evaluation and learning

    Dud \' k, M., Langford, J., and Li, L. Doubly robust policy evaluation and learning. In Proceedings of the 28th International Conference on Machine Learning, pp.\ 1097--1104, 2011

    work page 2011

  7. [7]

    On explore-then-commit strategies

    Garivier, A., Lattimore, T., and Kaufmann, E. On explore-then-commit strategies. Advances in Neural Information Processing Systems, 29, 2016

    work page 2016

  8. [8]

    A., Zhan, R., Wager, S., and Athey, S

    Hadad, V., Hirshberg, D. A., Zhan, R., Wager, S., and Athey, S. Confidence intervals for policy evaluation in adaptive experiments. Proceedings of the National Academy of Sciences, 118 0 (15): 0 e2014602118, 2021

    work page 2021

Show all 29 references
  1. [9]

    Hahn, J., Hirano, K., and Karlan, D. S. Adaptive experimental design using the propensity score. Journal of Business & Economic Statistics, 29: 0 108 -- 96, 2009

    2009

  2. [10]

    P., Thomas, P

    Hanna, J. P., Thomas, P. S., Stone, P., and Niekum, S. Data-efficient policy evaluation through behavior policy search. In International Conference on Machine Learning, 2017

    2017

  3. [11]

    R., Ramdas, A., McAuliffe, J

    Howard, S. R., Ramdas, A., McAuliffe, J. D., and Sekhon, J. S. Time-uniform, nonparametric, nonasymptotic confidence sequences. The Annals of Statistics, 2021

    2021

  4. [12]

    Off-policy risk assessment in contextual bandits

    Huang, A., Leqi, L., Lipton, Z., and Azizzadenesheli, K. Off-policy risk assessment in contextual bandits. Advances in Neural Information Processing Systems, 34: 0 23714--23726, 2021

    2021

  5. [13]

    Off-policy risk assessment for markov decision processes

    Huang, A., Leqi, L., Lipton, Z., and Azizzadenesheli, K. Off-policy risk assessment for markov decision processes. In International Conference on Artificial Intelligence and Statistics, pp.\ 5022--5050. PMLR, 2022

    2022

  6. [14]

    and Li, L

    Jiang, N. and Li, L. Doubly robust off-policy value evaluation for reinforcement learning. In International Conference on Machine Learning, pp.\ 652--661. PMLR, 2016

    2016

  7. [15]

    Efficient adaptive experimental design for average treatment effect estimation

    Kato, M., Ishihara, T., Honda, J., and Narita, Y. Efficient adaptive experimental design for average treatment effect estimation. arXiv preprint arXiv:2002.05308, 2020

    2002 arXiv

  8. [16]

    J., Denil, M., and de Freitas, N

    Konyushova, K., Chen, Y., Paine, T., Gulcehre, C., Paduraru, C., Mankowitz, D. J., Denil, M., and de Freitas, N. Active offline policy selection. Advances in Neural Information Processing Systems, 34: 0 24631--24644, 2021

    2021

  9. [17]

    and Szepesv \'a ri, C

    Lattimore, T. and Szepesv \'a ri, C. Bandit algorithms. Cambridge University Press, 2020

    2020

  10. [18]

    Unbiased offline evaluation of contextual-bandit-based news article recommendation algorithms

    Li, L., Chu, W., Langford, J., and Wang, X. Unbiased offline evaluation of contextual-bandit-based news article recommendation algorithms. In Proceedings of the fourth ACM international conference on Web search and data mining, pp.\ 297--306, 2011

    2011

  11. [19]

    Toward minimax off-policy value estimation

    Li, L., Munos, R., and Szepesv \'a ri, C. Toward minimax off-policy value estimation. In Artificial Intelligence and Statistics, pp.\ 608--616. PMLR, 2015

    2015

  12. [20]

    Optimal treatment allocation for efficient policy evaluation in sequential decision making

    Li, T., Shi, C., Wang, J., Zhou, F., et al. Optimal treatment allocation for efficient policy evaluation in sequential decision making. Advances in Neural Information Processing Systems, 36, 2024

    2024

  13. [21]

    Ma, C., Zhu, B., Jiao, J., and Wainwright, M. J. Minimax off-policy evaluation for multi-armed bandits. IEEE Transactions on Information Theory, 68: 0 5314--5339, 2021

    2021

  14. [22]

    Logarithmic N eyman regret for adaptive estimation of the average treatment effect

    Neopane, O., Ramdas, A., and Singh, A. Logarithmic N eyman regret for adaptive estimation of the average treatment effect. AISTATS, 2025

    2025

  15. [23]

    On the two different aspects of the representative method: the method of stratified sampling and the method of purposive selection

    Neyman, J. On the two different aspects of the representative method: the method of stratified sampling and the method of purposive selection. Journal of the Royal Statistical Society, 97: 0 123--150, 1934

    1934

  16. [24]

    Thompson, W. R. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25 0 (3-4): 0 285--294, 1933

    1933

  17. [25]

    Optimal and adaptive off-policy evaluation in contextual bandits

    Wang, Y.-X., Agarwal, A., and Dud k, M. Optimal and adaptive off-policy evaluation in contextual bandits. In International Conference on Machine Learning, pp.\ 3589--3597. PMLR, 2017

    2017

  18. [26]

    and Ramdas, A

    Waudby-Smith, I. and Ramdas, A. Estimating means of bounded random variables by betting. Journal of the Royal Statistical Society Series B Methodological, 2023

    2023

  19. [27]

    Anytime-valid off-policy inference for contextual bandits

    Waudby-Smith, I., Wu, L., Ramdas, A., Karampatziakis, N., and Mineiro, P. Anytime-valid off-policy inference for contextual bandits. ACM/JMS Journal of Data Science, 2022

    2022

  20. [28]

    Inference for batched bandits

    Zhang, K., Janson, L., and Murphy, S. Inference for batched bandits. Advances in Neural Information Processing Systems, 33: 0 9818--9829, 2020

    2020

  21. [29]

    Statistical inference with m-estimators on adaptively collected data

    Zhang, K., Janson, L., and Murphy, S. Statistical inference with m-estimators on adaptively collected data. Advances in Neural Information Processing Systems, 34: 0 7460--7471, 2021

    2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.