Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Revisiting Optimal Allocations for Binary Responses: Insights from Considering Type-I Error Rate Control

T0 review · 3 major / 5 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 Two new response-adaptive allocation proportions, derived from the score test with finite-sample estimators, keep type-I error rates near nominal levels in binary-outcome trials while preserving power and improving patient outcomes.

desk verdict A genuine practical finding—Wald-based optimal allocations severely inflate type-I error—and a mostly-working score-test fix, but the control claim is overstated by an uncharacterized simulation domain and boundary failures. read the letter →

arxiv 2502.06381 v4 pith:I5OD3T4T submitted 2025-02-10 stat.ME math.STstat.APstat.TH

Lukas Pin , Sofía S. Villar , William F. Rosenberger This is my paper · ORCID
classification stat.MEmath.STstat.APstat.TH MSC 62K0562L0562P10
keywords NeymanallocationRSHIRscoretestWaldpatientbenefitresponse-adaptiverandomizationtype-Ierrorratecontrolbinaryoutcomes
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

Classic optimal response-adaptive allocations for binary outcomes are derived for the Wald test, and this paper shows that using them with that test inflates the type-I error rate dramatically, in some null scenarios to over 80%, an inflation it says was not previously documented. Existing remedies from the literature, including adjusted variance estimators, equal-probability sampling when a variance estimate is zero, and longer burn-in periods, all fail to restore error control. The paper's solution is to re-derive the two optimal proportions for the score test rather than the Wald test, substituting finite-sample estimates for the unknown true success probabilities in the optimization. The resulting Neyman-like proportion $\rho_{N0}^{n}$ and RSHIR-like proportion $\rho_{R0}^{n}$ keep the rejection rate of the score test near the nominal 5% level across most of the parameter space in finite samples, with power within a few percent of complete randomization and increased expected successes, as illustrated on an early-phase and a confirmatory trial redesign. If correct, this gives practitioners a response-adaptive design that meets the regulatory minimum of type-I error control without giving up the patient-benefit gains of adaptation.

What carries the argument

The load-bearing object is the score test for equality of two proportions, $Z_0 = (\hat p_1-\hat p_0)/\sqrt{\hat p \hat q (1/n_0+1/n_1)}$, where $\hat p$ is the pooled success proportion; its square is the Pearson chi-square statistic. The paper rewrites the variance of $Z_0$ in terms of $\hat p_0$, $\hat p_1$, $n_0$, and $n_1$, and plugs the current maximum-likelihood estimates into the Neyman-type and RSHIR-type optimization problems that were originally formulated with unknown true probabilities and Wald-test variance. Minimizing this variance over the allocation fraction gives the closed-form $\rho_{N0}^{n}$; minimizing expected failures under the same variance constraint gives $\rho_{R0}^{n}$, which is solved numerically. The ERADE targeting rule then pushes the realized randomization probabilities toward these estimated per-step targets, so the whole procedure is an estimator-driven variant of the classical optimal-allocation designs built around the test whose null distribution is actually used for inference.

What would settle it

Re-run the proposed designs under the null at $n=50$ with $p_0=p_1=0.9$ and record how often the score test rejects at the nominal 5% level; the paper reports 7.9% for $\rho_{R0}^{n}$, already above nominal. Repeating this over the full $(p_0,p_1)$ grid, or computing the exact null distribution of $Z_0$ under the estimator-driven allocation, would establish whether the advertised error control extends to the boundary or holds only in the interior of the parameter space.

Watch

Extended reading notes

Core claim

The central claim is that an optimal response-adaptive allocation must be derived from the test statistic that will actually be used for inference, with unknown parameters replaced by their estimators, and that this is what preserves the type-I error rate. The classic Neyman and RSHIR proportions, $\rho_{N1}$ and $\rho_{R1}$, solve two variance-constrained optimization problems for the Wald test $Z_1$; the new proportions solve the same problems for the score test $Z_0$, whose variance is the pooled null variance. The power-maximizing proportion is $\rho_{N0}^{n} = \sqrt{\hat p_0 \hat q_0}/(\sqrt{\hat p_0 \hat q_0}+\sqrt{\hat p_1 \hat q_1})$, which is the complement of the estimated classic Neyman proportion, and the failure-minimizing proportion $\rho_{R0}^{n}$ is computed numerically from the score-test variance constraint. Driven by ERADE with a minimal burn-in of two patients per arm, these rules control the type-I error rate for the largest part of the parametric space, keep power comparable to complete randomization, and increase the expected number of successes; in the confirmatory trial redesign the RSHIR-like rule yields about 25 additional successes out of 1502 patients.

Load-bearing premise

The claim stands or falls on whether feeding estimated success probabilities into the allocation rule at each patient's turn keeps the final test statistic behaving like a standard normal random variable under the null; the paper's own simulation at $p_0=p_1=0.9$ shows this is not true near the boundary, where the type-I error reaches 7.9%.

Editorial extensions

If this is right

  • Used with the Wald test, the classical Neyman and RSHIR allocations can inflate the type-I error rate to more than 80% under the null, so those procedures are not reliable for confirmatory trials as originally proposed.
  • The new score-test proportions $\rho_{N0}^{n}$ and $\rho_{R0}^{n}$ hold the type-I error rate near 5% across most of the parametric space at the simulated sample sizes, with only two patients per arm of burn-in.
  • Power of the new designs stays within a few percent of complete randomization, while the RSHIR-like allocation increases the expected number of successes, adding about 25 successes in the confirmatory trial redesign.
  • Because the optimization uses finite-sample estimates rather than true parameters, the designs are implementable without prior knowledge of success probabilities, and the paper reports no meaningful increase in variability from estimation.
  • The same derivation principle extends to other effect measures such as relative risk or odds ratio, and to multi-armed trials, though multi-arm versions need constraints to prevent allocating only to extreme arms.

Reading between the lines

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

  • The complement relation $\rho_{N0}^{n}=1-\hat\rho_{N1}$ means the score-test Neyman rule sends more patients to the arm with the smaller estimated Bernoulli variance; when that arm is also the better arm, this is the mechanism behind the patient-benefit gain at roughly unchanged power.
  • The boundary inflation the paper reports at $p_0=p_1=0.9$ (7.9% type-I error for $\rho_{R0}^{n}$) suggests a practical safeguard the authors do not propose: constrain the estimated target proportions away from 0 and 1, or calibrate the score test by permutation in extreme-parameter regions, to extend control to the full grid.
  • Because the derivation only needs a null variance that is a function of a common parameter, a testable extension is to repeat the optimization for score tests of relative risk or odds ratio and check whether the resulting allocations inherit the same error-control pattern.
  • The confirmatory-trial gain suggests that in high-success-rate settings the RSHIR-like allocation can deliver a meaningful patient-benefit improvement at negligible power cost; evaluating it under group-sequential or delayed-outcome logistics, which the paper sets aside, is a natural next test.
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 / 5 minor

Summary. The paper revisits optimal response-adaptive randomization (RAR) proportions for two-armed trials with binary endpoints. It shows that Neyman and RSHIR allocations, when used with the Wald test, can produce severe type-I error inflation (Table 1), that several published fixes do not fully solve the problem (Section 3), and that deriving allocations for the score test with finite-sample estimators gives two new rules, rho_N0 and rho_R0, intended to control type-I error while preserving power and improving patient benefit. The new rules are derived by minimizing the estimated score-test variance (Equation 2) and are evaluated by simulation and by redesigning two published trials (Section 5).

Significance. If the validity domain can be made precise, this is a useful contribution: it documents a striking and practically important inflation of the Wald-based Neyman/RSHIR allocations, proposes a principled score-test-based modification, and supports it with reproducible simulations (10^4 replicates, stated Monte Carlo error) and two externally grounded trial redesigns. The central algebra in Section 4.1 is internally consistent, the public code is a strength, and the paper's framing that type-I error control should precede power or patient-benefit considerations is appropriate. The main limitation is that the central claim of 'controlling type-I error rate' is currently supported only by simulation on a grid, and the paper's own Table 4 shows the nominal level is exceeded near the boundary of the parameter space.

major comments (3)
  1. [Abstract; Section 6] The abstract and Section 6 claim that the proposed designs control the type-I error rate, but Table 4 shows the nominal 0.05 level is exceeded at several simulated points: for n=50 with burn-in 2, rho_R0 gives 7.9% at p0=p1=0.9 and 6.0% at p0=p1=0.1, while rho_N0 gives 6.8% at p0=p1=0.1 and 6.1% at p0=p1=0.9. The Section 6 statement that the designs control type-I error over '90% of the parametric space' is not derived and appears to be an informal summary of the simulated grid. The authors should either restrict the claim to a characterized region (for example, with a stated tolerance and p0,p1 bounded away from 0 and 1) or provide a formal guarantee.
  2. [Section 4.1; Section 4.3; Section 6] No argument is given that the final score statistic Z0 retains its standard normal null distribution when the allocation process is driven by the same accumulating responses that enter the pooled estimator. Minimizing the estimated variance in Equation (2) stepwise does not by itself establish that the unconditional null distribution of Z0 is unchanged under the ERADE assignment mechanism; the simulations in Table 4 are evidence rather than a proof, and the table shows boundary failures. I recommend either stating explicit conditions (target proportions bounded away from 0 and 1, a precise rule for zero estimated variances, and an ERADE convergence condition) with a proof or conservative bound, or changing the claim to approximate control within simulation tolerance on a specified grid.
  3. [Section 4.2; Table 4] The derivation of rho_R0 is not fully verifiable as printed: the objective and derivative formulas in Section 4.2 have unbalanced notation and are not accompanied by a uniqueness or global-minimum argument for the numerical solution. Since rho_R0 is one of the two proposed designs, the paper should specify the root-finding procedure (range, starting values, root selection) and confirm that the simulations use that exact rule. In addition, the handling of zero estimated variances under the new rules is not stated; with burn-in 2 and p near 0 or 1 this is not a purely theoretical concern.
minor comments (5)
  1. [Section 2.3] The text first says the simulation section used a minimal burn-in period of B=4 and later says 'We use a burn-in of 2 patients per arm consistently throughout the paper'; the captions and tables use 2 per arm, so the B=4 sentence appears to be an error and should be corrected for reproducibility.
  2. [Equation (2)] Equation (2) is typeset in a way that makes the algebra difficult to follow; the rewritten variance should be displayed with clear grouping of rho_n, 1-rho_n, and the cross term.
  3. [Section 4.2] The optimization problem for rho_R0 is missing explicit parentheses in the printed objective; the intended expression appears to be the product of expected failures and the score-test variance divided by C, but the displayed formula does not unambiguously show this.
  4. [Section 4.3; Table 4] The text says rho_R0 is 'slightly inflated for values around 0.8', but Table 4 reports 7.9% at p0=p1=0.9; the wording should match the table, and the boundary behavior should be described honestly.
  5. [Figure 1 caption] The caption lists 'the RSHIR proportion testing with the Wald test ρN1' but the correct symbol is ρR1; the duplicate ρN1 should be fixed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the new allocations are derived from the score-test variance and validated on external benchmarks, with no fitted quantity renamed as a prediction.

full rationale

The paper's central derivation chain starts from the score-test variance with the pooled estimator p-hat, rewrites it in terms of the allocation fraction rho_n (Eq. 2), and solves the resulting constrained optimization to obtain rho_N0 and rho_R0 (Sections 4.1-4.2). No parameter in these derivations is fitted to the type-I error or power outcomes that are later reported; the only inputs are the Bernoulli model, the score-test variance formula, and the declared optimization objectives. The resulting allocations are then evaluated by simulation and by redesigning two externally published trials (Kubo et al. 2023; Decousus et al. 2010), so the reported type-I error control and patient-benefit gains are not renamed fits. The identity rho_N0 = 1 - rho_N1 (Section 4.1) is a mathematical consequence of the two variance expressions, not a definition of the target outcome. Self-citations (Rosenberger et al. 2001, Hu and Rosenberger 2006, Pin et al. 2024, Tang et al. 2025) provide background, targeting procedures, or tables of known allocations; the new claim does not rest on an unverified self-cited theorem, and no uniqueness theorem is imported. The weakest premise — that the score statistic retains its null distribution under estimator-driven ERADE — is supported by simulation rather than proof, and the paper's own Table 4 shows boundary inflation (e.g., 7.9% at p0=p1=0.9 for rho_R0); however, this is an unproved condition and an overclaim in Section 6 ('90% of the parametric space'), not a circular reduction, so it does not raise the circularity score.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The new allocations introduce no fitted constants and no invented entities; the contribution is a redefined optimization objective matched to the score test. The claim's contingency rests on two hand-chosen design constants (ERADE alpha and burn-in) and four background assumptions, the sharpest being the ad hoc stepwise estimated-variance objective and the borrowed asymptotic theory for adaptive targeting. No quantities were fitted to outcome data to make the method work.

free parameters (2)
  • ERADE tuning constant alpha = 0.5
    Hand-chosen within Burman (1996)'s recommended 0.4-0.7 range (Section 2.2); controls allocation inertia and affects the simulated type-I error and lock-out dynamics.
  • Burn-in size B = 2 patients per arm (4 total)
    Deliberately minimal, chosen in Section 2.3 to probe maximal type-I error inflation; Table 2 shows operating characteristics are burn-in dependent, so the type-I error claims are contingent on this choice.
assumptions (4)
  • domain assumption Bernoulli i.i.d. responses per arm with immediate outcome observation and sequential single-patient accrual
    Trial model assumed throughout Section 2; delayed outcomes, group allocation, or early stopping would change the operating characteristics, and the paper acknowledges this only as a scope statement.
  • domain assumption Asymptotic normality of MLE-based test statistics under DBCD/ERADE targeting (Hu and Zhang 2004; Hu et al. 2009)
    Invoked implicitly by using a z-test at trial end; Section 2.2 cites the targeting procedures but no validity theorem is restated for the proposed estimator-driven target functions.
  • ad hoc to paper Minimizing the estimated score-test variance (Equation 2) stepwise is the right finite-sample proxy for score-test power
    The Section 4 optimization plugs p-hat values into the pooled variance; no argument shows stepwise estimated-variance minimization preserves the null distribution of Z0 in finite samples. This is the load-bearing modeling choice of the paper.
  • domain assumption Published success rates of the two redesigned trials are valid fixed scenarios
    Section 5 uses Kubo et al. (2023) (p0=0.635, p1=0.893) and Decousus et al. (2010) (p0=0.941, p1=0.991) as true effect sizes; these are single-study point estimates treated as fixed, with no sensitivity analysis shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Optimal Allocations for Binary Responses: Insights from Considering Type-I Error Rate Control." pith.science (2026). https://pith.science/paper/I5OD3T4T

@misc{pith2026250206381,
  author       = {Pith},
  title        = {Pith review of: Revisiting Optimal Allocations for Binary Responses: Insights from Considering Type-I Error Rate Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I5OD3T4T}},
  note         = {Machine review of arXiv:2502.06381}
}
read the original abstract

This work revisits optimal response-adaptive designs from a type-I error rate perspective, highlighting when and how much these allocations exacerbate type-I error rate inflation - an issue previously undocumented. We explore a range of approaches from the literature that can be applied to reduce type-I error rate inflation. However, we found that all of these approaches fail to give a robust solution to the problem. To address this, we derive two optimal allocation proportions, incorporating the more robust score test (instead of the Wald test) with finite sample estimators (instead of the unknown true values) in the formulation of the optimization problem. One proportion optimizes statistical power and the other minimizes the total number failures in a trial while maintaining a fixed variance level. Through simulations based on an early-phase and a confirmatory trial we provide crucial practical insight into how these new optimal proportion designs can offer substantial patient outcomes advantages while controlling type-I error rate. While we focused on binary outcomes, the framework offers valuable insights that naturally extend to other outcome types, multi-armed trials and alternative measures of interest.

Figures

Figures reproduced from arXiv: 2502.06381 by the authors.

Figure 1
Figure 1. Comparison of Type-I error rates of: the Neyman proportion testing with the Wald test [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Optimal proportions ρ n R0 , ρR1 , ρ n N0 , and ρN1 given p0 = 0.3, 0.5, 0.9 and p1 ∈ (0, 1). 16 [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 20 canonical work pages

  1. [6]

    doi:10.1186/s12874-021- 01303-z

    ISSN 1471-2288. doi:10.1186/s12874-021- 01303-z. 12 Revisiting Optimal Proportions for Binary Responses A PREPRINT A. Biswas, S. Mandal, and R. Bhattacharya. Multi-treatment optimal response-adaptive designs for phase III clinical trials. Journal of the Korean Statistical Society, 40:33–44,

    work page doi:10.1186/s12874-021-

  2. [12]

    doi:10.1007/978-3-319-18539-2

    ISBN 978-3-319-18538-5. doi:10.1007/978-3-319-18539-2. F. Hu and W. F. Rosenberger.The Theory of Response-Adaptive Randomization in Clinical Trials. Wiley,

    doi:10.1007/978-3-319-18539-2

  3. [14]

    T. P. Morris, I. R. White, and M. J. Crowther. Using simulation studies to evaluate statistical methods. Statistics in Medicine, 38:2074–2102,

    work page

  4. [21]

    URL https://doi.org/10.1007/978-3-031-65948-5_12

    doi:10.1007/978-3-031-65948-5_12. URL https://doi.org/10.1007/978-3-031-65948-5_12 . L. Pin, M. Neubauer, D. Robertson, and S. S. Villar. Response adaptive clinical trials. https://doi.org/10.5281/ zenodo.14691304,

    work page doi:10.1007/978-3-031-65948-5_12

  5. [22]

    doi:10.1214/22-STS865. W. F. Rosenberger and J. M. Lachin. Randomization in Clinical Trials . Wiley,

    doi:10.1214/22-sts865

  6. [25]

    doi:10.1007/s42519-019-0073-4. E. Y . N. Tang, S. Baas, D. Kaddaj, L. Pin, D. S. Robertson, and S. S. Villar. A burn-in(g) question: How long should an initial equal randomization stage be before bayesian response-adaptive randomization?,

    work page doi:10.1007/s42519-019-0073-4

  7. [29]

    URL https://doi.org/10.1093/biomtc/ujae049

    doi:10.1093/biomtc/ujae049. URL https://doi.org/10.1093/biomtc/ujae049. ujae049. H. Zhu and F. Hu. Implementing optimal allocation for sequential continuous responses with multiple treatments. Journal of Statistical Planning and Inference, 139:2420–2430,

    work page doi:10.1093/biomtc/ujae049

  8. [30]

    doi:10.1016/j.jspi.2008.11.003

    ISSN 03783758. doi:10.1016/j.jspi.2008.11.003. 14 Revisiting Optimal Proportions for Binary Responses A PREPRINT 0.00 0.25 0.50 0.75 0.00 0.25 0.50 0.75 1.00 p1 Type−I Error Rate Method ρ N1 ρ R1 n=50 0.00 0.02 0.04 0.06 0.08 0.00 0.25 0.50 0.75 1.00 p1 Type−I Error Rate Method ρ CR1 ρ CR0 ρ N0 n ρ R0 n n=50 Figure 1: Comparison of Type-I error rates of: ...

    work page doi:10.1016/j.jspi.2008.11.003 2008

Show all 30 references
  1. [1933]

    doi:10.1093/biomet/25.3-4.285

    ISSN 0006-3444. doi:10.1093/biomet/25.3-4.285. A. A. Tschuprow. On the mathematical expectation of the moments of frequency distributions in the case of correlated observations. Metron, 2:461–493, 646–683,

    doi

  2. [1934]

    doi:10.2307/2342192

    ISSN 09528385. doi:10.2307/2342192. L. Pin, S. S. Villar, and W. F. Rosenberger. Response-adaptive randomization designs based on optimal allocation proportions. In D. G. Chen, editor, Biostatistics in Biopharmaceutical Research and Development. Springer,

    doi

  3. [1977]

    doi:10.1080/00031305.1977.10479225. J. R. Eisele. The doubly adaptive biased coin design for sequential clinical trials. Journal of Statistical Planning and Inference, 38:249–261,

    1977 arXiv

  4. [1994]

    doi:10.1016/0378-3758(94)90038-8

    ISSN 03783758. doi:10.1016/0378-3758(94)90038-8. J. R. Eisele and M. B. Woodroofe. Central limit theorems for doubly adaptive biased coin designs. The Annals of Statistics, 23:234–254,

    doi

  5. [2000]

    doi:10.1080/00031305.2000.10474560. D. Azriel and P. D. Feigin. Adaptive designs to maximize power in clinical trials with multiple treatments. Sequential Analysis, 33:60–86,

    2000 arXiv

  6. [2001]

    doi:10.1111/j.0006-341X.2001.00909.x

    ISSN 0006341X. doi:10.1111/j.0006-341X.2001.00909.x. 13 Revisiting Optimal Proportions for Binary Responses A PREPRINT D. W. Scott. Statistics: A Concise Mathematical Introduction for Students, Scientists, and Engineers. John Wiley & Sons,

    2001 arXiv

  7. [2002]

    doi:10.1002/0471249688

    ISBN 9780471360933. doi:10.1002/0471249688. A. Agresti and B. Caffo. Simple and effective confidence intervals for proportions and differences of propor- tions result from adding two successes and two failures. The American Statistician , 54(4):280–288,

    doi

  8. [2004]

    doi:10.1214/aos/1079120137

    ISSN 0090-5364. doi:10.1214/aos/1079120137. F. Hu, L. X. Zhang, and X. He. Efficient randomized-adaptive designs. Annals of Statistics, 37:2543–2560,

  9. [2006]

    doi:10.1002/047005588X

    ISBN 9780471653967. doi:10.1002/047005588X. F. Hu and L.-X. Zhang. Asymptotic properties of doubly adaptive biased coin designs for multitreatment clinical trials. The Annals of Statistics, 32:268–301,

    doi

  10. [2007]

    doi:10.1198/016214506000000906. X. Ye, F. Hu, and W. Ma. Robustness of response-adaptive randomization. Biometrics, 80(2), June

    doi

  11. [2009]

    doi:10.1214/08-AOS655

    ISSN 00905364. doi:10.1214/08-AOS655. C. Jennison and B. W. Turnbull. Group Sequential Methods with Applications to Clinical Trials. CRC Press,

    doi

  12. [2010]

    doi:10.1198/sbr.2009.0056

    ISSN 1946-6315. doi:10.1198/sbr.2009.0056. Y . Kubo, T. Makino, M. Yamasaki, K. Tanaka, K. Yamashita, O. Shiraishi, et al. Three-course neoadjuvant chemotherapy associated with unfavorable survival of non-responders to the first two courses for locally advanced esophageal canc...

    1946

  13. [2011]

    doi:10.1016/j.jkss.2010.04.004

    ISSN 12263192. doi:10.1016/j.jkss.2010.04.004. C.-F. Burman. On sequential treatment allocations in clinical trials (doctoral dissertation)

    2010 doi

  14. [2014]

    doi:10.1080/07474946.2014.856637

    ISSN 0747-4946. doi:10.1080/07474946.2014.856637. D. Azriel, M. Mandel, and Y . Rinott. Optimal allocation to maximize the power of two-sample tests for binary response. Biometrika, 99(1):101–113,

    2014

  15. [2015]

    URL https://doi.org/10.1186/s13063-015-0958-9

    doi:10.1186/s13063-015- 0958-9. URL https://doi.org/10.1186/s13063-015-0958-9 . K. R. Eberhardt and M. A. Fligner. A comparison of two tests for equality of two proportions.The American Statistician, 31(4):151–155,

    doi

  16. [2016]

    doi:10.1002/9781118742112

    ISBN 9781118742242. doi:10.1002/9781118742112. W. F. Rosenberger, N. Stallard, A. Ivanova, C. N. Harper, and M. L. Ricks. Optimal adaptive designs for binary response trials. Biometrics, 57:909–913,

    doi

  17. [2019]

    doi:10.1002/sim.8086

    ISSN 0277-6715. doi:10.1002/sim.8086. J. Neyman. 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, 97:558,

    doi

  18. [2020]

    doi:10.1186/s12916- 020-01808-2. H. Decousus, P. Prandoni, P. Mismetti, R. M. Bauersachs, Z. Boda, B. Brenner, et al. Fondaparinux for the treatment of superficial-vein thrombosis in the legs. New England Journal of Medicine, 363:1222–1232,

    doi

  19. [2021]

    URL https://doi.org/10.1214/21-EJS1864

    doi:10.1214/21-EJS1864. URL https://doi.org/10.1214/21-EJS1864. V . W. Berger, L. J. Bour, K. Carter, J. J. Chipman, C. C. Everett, N. Heussen, et al. A roadmap to using randomization in clinical trials. BMC Medical Research Methodology, 21:168,

    doi

  20. [2023]

    Epub 2023 Jun

    doi:10.1245/s10434-023-13548-7. Epub 2023 Jun

    2023 doi

  21. [2024]

    URL https://arxiv.org/abs/2401.15694. A. Baldi Antognini, A. Vagheggini, and M. Zagoraiou. Is the classical wald test always suitable under response-adaptive randomization? Statistical Methods in Medical Research, 27(8):2294–2311,

    arXiv

  22. [2025]

    URL https://arxiv.org/abs/2503.19795. W. R. Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25:285–294,

Pith tools

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