Pith. sign in

REVIEW 3 major objections 3 minor 19 references

Adaptive Experimental Design Using Shrinkage Estimators

T0 review · 3 major / 3 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Adaptive trials can be steered to minimize the risk of a shrinkage estimator, not just the variance of a difference-in-means estimator.

desk verdict New idea worth refereeing: minimizing shrinkage-estimator risk in adaptive trials via Gaussian quadratic-form integrals, but the adaptive extension rests on an unproved conditional-unbiasedness assumption. read the letter →

arxiv 2602.07404 v2 pith:25W3QXJE submitted 2026-02-07 stat.ME math.STstat.TH

Evan T. R. Rosenman , Kristen B. Hunter This is my paper
classification stat.MEmath.STstat.TH MSC 62K0562L05
keywords adaptiveexperimentaldesignshrinkageestimationStein'sunbiasedriskestimatemulti-armedtrialsNeymanallocationGaussianquadraticformssequentialtreatmenteffect
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

This paper argues that adaptive treatment allocation in multi-armed trials should target the risk of a shrinkage estimator rather than the variance of unbiased difference-in-means estimates. It shows that for three Stein-type estimators (Bock, SURE-minimizing, and Dimmery), the expected squared-error loss can be written as a linear combination of expectations of ratios of Gaussian quadratic forms, and that these expectations reduce to one-dimensional integrals that can be evaluated quickly. This makes a practical sequential design possible: after a short burn-in, each new arrival is assigned to whichever arm most reduces the current estimated shrinker risk. The paper further claims that risk-minimizing allocations put more units on the control arm than Neyman allocation, and that simulations show meaningful mean-squared-error reductions over complete randomization and Neyman allocation, especially when signal-to-noise is low.

What carries the argument

The engine is Stein's Unbiased Risk Estimate (SURE) combined with an integral representation for expectations of ratios of quadratic forms in Gaussian vectors. SURE rewrites the risk of any estimator of the form hat-tau minus g(hat-tau) as tr(Sigma) plus a tractable expectation involving g and its Jacobian; for Bock, SURE-min, and Dimmery estimators, the leftover terms are ratios like E[||hat-tau||^2 / (hat-tau^T Sigma^{-1} hat-tau)^2]. Standardizing hat-tau to a standard normal vector turns each such ratio into a one-dimensional integral that can be evaluated numerically in milliseconds, which is what makes online risk-minimizing allocation feasible.

What would settle it

Run a simulated K=6 trial where the first burn-in batch contains a large chance imbalance, record the empirical distribution of hat-tau across repetitions under the greedy rule, and compare its mean and covariance with the assumed Gaussian with plug-in variances; if the covariance is systematically larger or the distribution is skewed at moderate n_k, the integral-based risk estimates and the allocation decisions based on them will be mis-calibrated.

Watch

Extended reading notes

Core claim

The central claim is that the expected squared-error risk of a broad class of shrinkage estimators for a vector of treatment effects can be expressed as a sum of expectations of ratios of Gaussian quadratic forms, and those expectations can be computed accurately and quickly by one-dimensional numerical integration. The authors establish this formally for Bock's estimator and apply the same representation to the SURE-min and Dimmery estimators. Under the maintained approximation that the difference-in-means estimator is Gaussian with covariance Sigma, this turns the design problem—choosing arm sample sizes to minimize shrinker loss—into an optimizable, computable objective. The authors use t

Load-bearing premise

The whole risk computation and the greedy assignment rule rely on the difference-in-means estimator remaining centered Gaussian with covariance diag(V_k/n_k) + (V_0/n_0)11^T even though assignments are chosen deterministically from past data; the paper assumes adaptivity is not so aggressive as to invalidate this, without giving explicit conditions.

Editorial extensions

If this is right

  • A trial platform can, after a burn-in, send each new arrival to the arm that minimizes the estimated shrinker risk using only running means, variances, and a fast integral evaluation.
  • If the paper's simulations hold, this greedy allocation can cut estimation error substantially relative to complete randomization and sequential Neyman allocation in low-signal, high-noise settings.
  • Risk-minimizing designs are characterizable in advance: they allocate more units to control than Neyman, and for SURE-min and Dimmery also toward high-variance arms; sparsity of the true effects changes the allocation, especially for Dimmery.
  • The domination conditions in Lemmas 2 and 3 give practitioners a post-trial check: if the covariance satisfies the stated trace and eigenvalue inequalities, the shrinker is guaranteed lower expected squared error than the difference-in-means estimator.
  • The computational machinery is not tied to the three estimators; any shrinker whose risk is a ratio of Gaussian quadratic forms can be plugged into the same adaptive algorithm.

Reading between the lines

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

  • The paper leaves open whether the Gaussian approximation survives under a deterministic greedy rule; a natural extension is to randomize assignment probabilities and re-derive the risk integrals under the resulting asymptotic distribution—if normality still holds, the same machinery applies.
  • The consistent finding that control-arm oversampling helps shrinkers suggests a cheap heuristic: an analyst could pre-specify an elevated control proportion rather than run full adaptivity and expect most of the benefit at low signal.
  • Dimmery's overweighting of a sparse strong-signal arm hints that shrinker-risk allocation interpolates between pure estimation efficiency and best-arm identification; one could test whether the greedy rule's allocations track Thompson-sampling probabilities when the true effects are sparse.
  • The fast integral evaluation also opens a design-calibration tool: before launching a trial, simulate candidate allocations and compute exact shrinker risk for guessed parameters, which could replace asymptotic heuristics for choosing burn-in length.
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

3 major / 3 minor

Summary. The paper studies sequential multi-armed trials with K active treatments plus a control, where the final causal estimates are obtained by Stein-type shrinkage estimators (Bock, SURE-min, and Dimmery). It derives SURE-based risk expressions and domination conditions for each shrinker, characterizes oracle risk-minimizing allocations in a low signal-to-noise regime (concluding that those allocations put more mass on control than the Neyman allocation), and then proposes a greedy adaptive algorithm: at each arrival after a burn-in, assign the next unit to the arm that minimizes the current plug-in estimate of the shrinker risk, with the risk computed by numerical integration of ratios of Gaussian quadratic forms. Simulations compare adaptive risk-minimizing allocation with complete randomization and Neyman allocation, reporting MSE gains for SURE-min and Dimmery at low signal-to-noise ratios.

Significance. If the underlying assumptions hold, the paper offers a useful and practical way to target the risk of a shrinkage estimator rather than the variance of an unbiased estimator, and the one-dimensional integral representation for risk is a genuine computational contribution. The SURE derivations and domination lemmas in Section 3 and the appendices are careful and give checkable conditions for practitioners. The simulation evidence in Table 4 suggests large potential gains, especially in low-signal, high-variance regimes. However, the central extension to adaptivity rests on an unproved normality assumption under the deterministic greedy rule, and the exact integral formulas are only written out for Bock's estimator. These gaps currently limit the manuscript's claims.

major comments (3)
  1. [Section 3.2 and 5.2] The Gaussian model τ̂ ~ N(τ, Σ) is the foundation of the risk formulas used in the adaptive rule, but for the greedy rule it is asserted rather than derived. Under W_i = argmin_k R_δ(n(i-1)+e_k, ...), the realized arm sizes n_k are data-dependent. Conditioning on n, units in an arm are not an iid sample from the arm's potential-outcome distribution; even a two-period rule that assigns the second unit to the arm with the higher first-period outcome makes the final sample mean in the chosen arm upwardly biased. Hence E[τ̂|n] ≠ τ, and Eq. (16)-(17) compute a proxy rather than the actual MSE being minimized. The statement 'we explicitly assume adaptivity is not so aggressive as to undermine the CLT' (Section 3.2) is not a condition linking the deterministic argmin rule to normality. Please provide formal conditions (for example, vanishing allocation fractions, batched updates, or a martingal
  2. [Section 5.1 and Table 3] The efficient one-dimensional integral representation is explicitly derived only for Bock's estimator in Eq. (17). For SURE-min and Dimmery, the text states that the risk is also a linear combination of ratios of Gaussian quadratic forms and Table 3 reports timings, but no corresponding Bao-Kan integral formulas or code are provided. The adaptive rule in Section 5.2 is said to use the 'efficient RCPP integral representations' for all three estimators. Without explicit formulas for E(τ̂ᵀΣτ̂/‖τ̂‖⁴) and E(1/‖τ̂‖²) in the general case, or a supplementary derivation, the algorithm for SURE-min and Dimmery is not reproducible. Please add the integral expressions (or state that the implementation actually uses the Section 4.2 heuristic approximations instead).
  3. [Section 4.2.2/4.2.3 and Section 5.2] Even if the exact integrals were supplied, the plug-in risk R_δ(n(i-1)+e_k, V̂(i-1), τ̂(i-1)) conditions on the data-dependent sample sizes n(i-1). The risk identity in Eq. (16) is an unconditional expectation over τ̂ with mean τ and covariance Σ. Under the adaptive rule, the conditional distribution of τ̂ given n(i-1) is not N(τ, Σ), and the plug-in substitution is not justified. The simulation section should at least report the realized bias or the empirical coverage of the normality assumption for the final sample sizes; otherwise the performance gains in Table 4 may be an artifact of the mismatch between the proxy used by the algorithm and the true risk.
minor comments (3)
  1. [Table 4] The first row of Table 4 is unreadable: '36.903.436.20' and '20.942.143.48' appear to be missing separators. Please reformat the table.
  2. [Appendix G] Appendix G contains the incomplete sentence 'TODO explain Neyman.' This must be completed before publication.
  3. [References / Typesetting] There are several typographical issues: 'SUTV A' instead of 'SUTVA'; 'a a tradeoff' in the introduction; the reference 'Rosenzweig and Offer-Westort, 0' has year 0; and 'Fran¸ cois' has an encoding artifact. Please proofread carefully.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: risk formulas come from external SURE/Bao-Kan results, the adaptive rule is a design choice, and simulation MSEs are not fitted predictions.

full rationale

The paper's derivation chain is not circular. The risk expressions in Section 5.1 (Eqs. 16-17) are obtained by applying Stein's Unbiased Risk Estimate (Lemma 1) to the candidate shrinkers, then transforming Gaussian quadratic forms via the external, citable results of Bao and Kan (2013). The dominance lemmas (Lemmas 2 and 3) are proved from SURE plus stated eigenvalue conditions, not from the paper's own conclusions. The oracle analysis in Section 4 uses explicit approximations (Satterthwaite moment matching, proportionality assumptions, low-SNR expansions) that are disclosed as approximations rather than passed off as exact predictions. The adaptive greedy rule in Section 5.2 is a proposed assignment mechanism, not a fitted parameter being fed back in as a prediction; the reported MSE reductions in Table 4 come from simulations, not from reusing the paper's own risk formulas as an outcome. The only load-bearing distributional assumption is in Section 3.2, where the paper says: 'we explicitly assume this is not the case, and suppose adaptivity is not so aggressive as to undermine the CLT. Thus, approximately, hat-tau ~ N(tau, Sigma).' That is an explicit, acknowledged assumption, and if it fails the risk formulas become misspecified; but an unsupported assumption is a correctness/validity concern, not a circular reduction of the derivations to their inputs. Appendix G contains a 'TODO explain Neyman' placeholder, but that is an omitted expositional detail, not a circular step. No self-citation chain is load-bearing: the invariance/eigenvalue facts are cited to Bunch et al. (1978) and Strawderman (2003), and the ratio-of-quadratic-forms tool is cited to Bao and Kan (2013), none of which are the present authors' prior work. Accordingly, there is no step in which an input is defined in terms of an output, a fitted quantity is renamed a prediction, or a conclusion is forced by the authors' own citations.

Assumptions & free parameters 4 free parameters · 8 assumptions · 0 invented entities

The method introduces no new particles, forces, or entities and no fitted constants. The free parameters are simulation and algorithm design choices. The load-bearing unverified premise is the CLT under greedy adaptivity; the oracle and heuristic sections add stated approximation axioms but do not affect the main numerical-integration claim.

free parameters (4)
  • burn-in size N_burn / b1 = not specified
    Required by the greedy algorithm to obtain initial estimates of \tau and V; no default or guidance is given. Appendix G defines b1 but never sets it.
  • simulation iterations n.iters = not specified
    All MSE summaries average over iterations, but n.iters is never fixed. The precision of Table 4 and the trajectory plots cannot be assessed.
  • potential-outcome variance DGP parameters = log(350) and log-sd 0.60; V0 = 0.5 or 4 times average Vk
    Chosen by hand for the simulations. They determine which regimes are explored but are not part of the mathematical claim.
  • signal-to-noise grid \kappa = 0, 3, 9
    Chosen to represent no/low/moderate/high SNR. The low-SNR proofs require \kappa \le \epsilon for an unspecified small \epsilon.
assumptions (8)
  • domain assumption SUTVA and i.i.d. sampling from a super-population of potential outcomes
    Section 3.1. Standard causal-inference framework; if violated, difference-in-means estimates a different quantity.
  • standard math Bounded potential outcomes and a CLT for \hat\tau
    Section 3.2 invokes a CLT to justify \hat\tau approximately N(\tau,\Sigma).
  • ad hoc to paper The CLT continues to hold under the paper's greedy deterministic adaptive allocation
    Section 3.2: 'suppose adaptivity is not so aggressive as to undermine the CLT.' No condition links the greedy argmin rule to this premise; Hadad et al. (2021) is cited as a warning but not resolved.
  • domain assumption Oracle knowledge of V0,...,VK in Section 4
    Section 4 assumes potential-outcome variances are known a priori to derive Neyman and risk-minimizing allocations. This is relaxed in Section 5 but used to build intuition.
  • standard math Bao and Kan (2013) moments of ratios of Gaussian quadratic forms
    Section 5.1, Equation (17). External mathematical result used to convert risk expectations to univariate integrals.
  • standard math Satterthwaite moment-matching approximation for ||\hat\tau||^2
    Section 4.2.2. Standard approximation; used only for the heuristic SURE-min allocation analysis, not for the main adaptive algorithm.
  • ad hoc to paper Proportionality approximation \hat\tau^T \Sigma \hat\tau \approx [E(\hat\tau^T\Sigma\hat\tau)/E(||\hat\tau||^2)] ||\hat\tau||^2
    Section 4.2.2. Exact only when \Sigma is proportional to identity; used to derive the SURE-min heuristic.
  • ad hoc to paper Low-SNR approximations E(1/||\hat\tau||^2) \approx 1/tr(\Sigma) and E(\hat\tau^T A \hat\tau / ||\hat\tau||^4) \approx tr(A\Sigma)/tr(\Sigma)^2
    Section 4.2.3. Used for Dimmery's estimator heuristic oracle analysis; these are stated approximations, not proven bounds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Experimental Design Using Shrinkage Estimators." pith.science (2026). https://pith.science/paper/25W3QXJE

@misc{pith2026260207404,
  author       = {Pith},
  title        = {Pith review of: Adaptive Experimental Design Using Shrinkage Estimators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/25W3QXJE}},
  note         = {Machine review of arXiv:2602.07404}
}
read the original abstract

In multi-armed trials, adaptive designs are a popular way to increase estimation efficiency or identify optimal treatments. Several recent papers have proposed adaptive variants of the classical Neyman allocation to assign treatments in sequential trials, with the goal of minimizing the error of a Horvitz-Thompson-style estimator. However, this approach may be inefficient, because it fails to borrow information across the treatment arms. In this paper, we consider adaptivity in a sequential trial with K active treatments and a control, and suggest the use of Stein-like shrinkage estimators to obtain the final causal estimates. These estimators share information across arms, yielding provable reductions in expected squared error loss relative to estimating each causal effect in isolation. Moreover, for each of our candidate shrinkers, the risk is the expectation of ratios of Gaussian quadratic forms, and can be computed efficiently via numerical integration. Hence, we suggest a simple algorithm for sequential adaptivity: assign treatments to each new arrival by choosing the arm that will minimize the estimated shrinker loss. Through simulations, we demonstrate that this approach can yield meaningful reductions in estimation error, especially in the low signal-to-noise regime. We also characterize how our adaptive algorithm assigns treatments differently than would a sequential Neyman allocation, and suggest a method for constructing shorter confidence intervals at the trial's conclusion.

Figures

Figures reproduced from arXiv: 2602.07404 by the authors.

Figure 1
Figure 1. Oracle allocations for K = 12 and N = 1, 000 with low control variance, and κ = 0. All shrinker￾risk-minimizing designs shift mass toward the control arm relative to the Neyman allocation, with the largest control shift under Bock’s estimator δˆB. SURE-min δˆS and Dimmery δˆD exhibit the same trend but with smaller magnitude. Crucially, the integrals in (17) are univariate, and hence can be efficiently computed to h… view at source ↗
Figure 2
Figure 2. Oracle allocations for K = 12 and N = 1, 000 with low control variance, κ = 9, and dense τ . Compared to the κ = 0 case ( [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Oracle allocations for K = 12 and N = 1, 000 with low control-variance regime, κ = 9, and sparse τ . Here, we see two key trends. First, the control reallocation for Bock’s estimator is much larger than in the otherwise-identical case in which τ is dense ( [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: κ = 0 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: τ sparse, κ = 3 [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: τ sparse, κ = 9 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: τ dense, κ = 3 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: τ dense, κ = 9. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Low-variance control regime, κ = 0 [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]
Figure 10
Figure 10. Figure 10: Low-variance control regime, τ sparse, κ = 3 [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]
Figure 11
Figure 11. Figure 11: Low-variance control regime, τ sparse, κ = 9. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_11.png]
Figure 12
Figure 12. Figure 12: Low-variance control regime, τ dense, κ = 3 [PITH_FULL_IMAGE:figures/full_fig_p029_12.png]
Figure 13
Figure 13. Figure 13: Low-variance control regime, τ dense, κ = 9. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 5 canonical work pages

  1. [1]

    ∂˜p ∂n0 = KV 0 n2 0 tr(Σ) K (1⊤ν NA)2 −λ max(Σ) λmax(Σ)2 and ∂˜p ∂nk = Vk n2 k tr(Σ)ν 2 k −λ max(Σ) λmax(Σ)2 , k= 1,

    We begin by computing the relevant derivatives at the Neyman allocation. ∂˜p ∂n0 = KV 0 n2 0 tr(Σ) K (1⊤ν NA)2 −λ max(Σ) λmax(Σ)2 and ∂˜p ∂nk = Vk n2 k tr(Σ)ν 2 k −λ max(Σ) λmax(Σ)2 , k= 1, . . . K. We also know that, at the Neyman allocation, KV 0 n2 0 = V1 n2 1 =· · ·=VK n2 K (see the Appendix, Section D for more details). Hence, the derivative inn 0 wi...

    1978

  2. [2]

    Applying independence and the prior result, we obtain E ∥ˆτ∥2 2 (ˆτ⊤Σ−1 ˆτ)2 = tr(Σ) K(K−2)

    By symmetry,E(U ⊤ΣU) = tr(Σ)/K. Applying independence and the prior result, we obtain E ∥ˆτ∥2 2 (ˆτ⊤Σ−1 ˆτ)2 = tr(Σ) K(K−2) . Plugging these identities into Equation 19 yields the given risk. F Proof of Lemma 6 Proof.Denote asΣ NA the covariance matrix ofˆτat the Neyman allocation, which is given by ΣNA =S NA diag √ V1, . . . , √ VK + r V0 K 11 ⊤ ! ,(20) ...

    1978

  3. [6]

    URL https://doi.org/10.1515/jci-2023-0068

    doi: doi:10.1515/jci-2023-0068. URL https://doi.org/10.1515/jci-2023-0068. K.-C. Li et al. From Stein’s unbiased risk estimates to the method of generalized cross validation.The Annals of Statistics, 13(4):1352–1377,

    work page doi:10.1515/jci-2023-0068 2023

  4. [9]

    doi: 10.1111/rssb.12491. W. F. Rosenberger, N. Stallard, A. Ivanova, C. N. Harper, and M. L. Ricks. Optimal adaptive designs for binary response trials.Biometrics, 57(3):909–913,

    work page doi:10.1111/rssb.12491

  5. [10]

    URLhttps://doi.org/10.1086/735504

    doi: 10.1086/735504. URLhttps://doi.org/10.1086/735504. D. B. Rubin. Estimating causal effects of treatments in randomized and nonrandomized studies.Journal of Educational Psychology, 66(5):688,

    work page doi:10.1086/735504

  6. [1933]

    doi: 10.2307/2332286. L. Trippa, E. Q. Lee, P. Y. Wen, T. T. Batchelor, T. Cloughesy, G. Parmigiani, and B. M. Alexander. Bayesian adaptive randomized trial design for patients with recurrent glioblastoma. Journal of Clinical Oncology, 30(26):3258–3263,

    doi:10.2307/2332286

  7. [1934]

    doi: 10.2307/2342192. M. Offer-Westort, E. H. Kennedy, and L. Miratrix. Adaptive experiments with control augmentation. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 83(5):1053–1079,

    doi:10.2307/2342192

  8. [1979]

    doi: 10.1214/aos/1176344646. M. Blackwell, N. E. Pashley, and D. Valentino. Batch adaptive designs to improve effi- ciency in social science experiments. Working paper, Harvard University,

    arXiv

Show all 19 references
  1. [1982]

    doi: 10.1093/biomet/69.1.61. Y. Bao and R. Kan. On the moments of ratios of quadratic forms in normal random variables.Journal of Multivariate Analysis, 117:229–245,

    doi

  2. [1985]

    Maharaj, R

    A. Maharaj, R. Sinha, D. Arbour, I. Waudby-Smith, S. Z. Liu, M. Sinha, R. Addanki, A. Ramdas, M. Garg, and V. Swaminathan. Anytime-valid confidence sequences in an enterprise a/b testing platform. InCompanion Proceedings of the ACM Web Conference 2023, pages 396–400,

    2023

  3. [2003]

    Tabord-Meehan

    M. Tabord-Meehan. Stratification trees for adaptive randomization in randomized controlled trials.arXiv preprint arXiv:1806.05127,

    arXiv

  4. [2011]

    doi: 10.18637/jss.v040.i08. B. Efron.Large-scale inference: empirical Bayes methods for estimation, testing, and prediction. Cam- bridge University Press,

    doi

  5. [2012]

    URL https://doi.org/10.1200/JCO.2011.39.8420

    doi: 10.1200/JCO.2011.39.8420. URL https://doi.org/10.1200/JCO.2011.39.8420. PMID: 22649140. S. S. Villar, J. Bowden, and J. Wason. Multi-armed bandit models for the optimal design of clinical trials: benefits and challenges.Statistical science: a review journal of the Institu...

    2011 doi

  6. [2014]

    URL https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.6086

    doi: https://doi.org/10.1002/sim.6086. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.6086. X. Xie, S. Kou, and L. D. Brown. SURE estimates for a heteroscedastic hierarchical model.Journal of the American Statistical Association, 107(500):1465–1479, 2012a. Y. Xie, J. ...

    doi

  7. [2018]

    doi: 10.1561/2200000070. F. E. Satterthwaite. An approximate distribution of estimates of variance components.Biometrics bulletin, 2(6):110–114,

    doi

  8. [2020]

    J. Zhao. Adaptive neyman allocation.arXiv preprint arXiv:2309.08808,

  9. [2021]

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

    2002 arXiv

  10. [2023]

    BecauseΣ is full-rank and symmetric, there exists a unique orthogonal matrixQthat satisfies QΣQT =D, 22 forD= diag (d 1,

    A Proof of Lemma 1 Proof.The proof draws heavily from Lemma 4.1 and Theorem 4.1 in Strawderman (2003). BecauseΣ is full-rank and symmetric, there exists a unique orthogonal matrixQthat satisfies QΣQT =D, 22 forD= diag (d 1, . . . , dK ) a diagonal matrix. DefineZ=QX(and, equiv...

    2003

  11. [2024]

    F. Chen, S. Ge, J. Qian, and C. Harshaw. Sigmoid-ftrl: Design-based adaptive neyman allocation for aipw estimators.arXiv preprint arXiv:2511.19905,

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.