General characterization of asymptotic normality for estimation after a group sequential trial

After a group sequential trial — a study with pre-planned interim analyses at which the trial may be stopped early — the final sample size is random and depends on the observed data. Berckmoes, Ivanova, and Molenberghs proved that the ordinary sample mean collected at the random stopping time is asymptotically normal, in total variation, as the maximal sample size grows, in a broad but incomplete list of regimes covering many realistic boundary choices, and they showed by counterexample that the conditions cannot simply be dropped. The open problem is to complete this picture: within the source model, give necessary and sufficient conditions on the boundary-family parameter γ, the true mean µ, and the boundary function ψ under which the randomly stopped sample mean retains the asymptotic behavior it would have with a deterministic sample size, identify the actual limiting distributions where normality fails, and thereby delimit exactly when naive post-trial confidence intervals are and are not valid.

Background

A group sequential trial (GST) is a study design, dominant in modern clinical trials, in which the accumulating data are inspected at a fixed schedule of interim analyses and a pre-specified stopping rule terminates the trial early — for efficacy, futility, or harm — once a boundary is crossed. Because the stopping decision depends on the data, the number of observations finally collected, $N$ , is a random variable taking values in a finite pre-specified set. A long-standing concern, dating back at least to the 1980s, is that this data-dependent sample size biases and distorts estimation performed after the trial: Milanzi, Molenberghs, Alonso, Kenward, Tsiatis, Davidian, and Verbeke (Statistics in Biosciences, 2015) reviewed the resulting inferential complications and argued that, despite them, the ordinary sample average remains a viable estimator after a GST for a wide class of stopping rules — without, however, developing its asymptotic distribution theory.

The sharpest form of the question concerns asymptotic normality. In the fixed-sample setting, the sample mean of i.i.d. observations is asymptotically normal by the central limit theorem, and confidence intervals are built on that fact. For a randomly stopped mean, the classical tool is Anscombe's theorem, a random-index central limit theorem stating that a sum indexed by a random sample size is asymptotically normal provided the index grows to infinity in probability and the partial-sum process satisfies a uniform-continuity-in-probability condition; Berckmoes (Applied Categorical Structures, 2018) gave an abstract, 'approach-theoretic' version of this theorem (approach theory being a framework unifying metric and topological notions of convergence) and applied it to justify naive confidence intervals after a GST. The concrete analysis was carried out by Berckmoes, Ivanova, and Molenberghs in two papers. First (Computational Statistics & Data Analysis, 2018), for i.i.d. normal outcomes and a generic stopping rule based on the partial sums at the interim looks, they derived explicit formulas for the expected trial length, the bias, and the mean squared error of the sample mean, showed the bias and MSE vanish at optimal rates when the first look is not too early, and established asymptotic normality in total variation distance (a strong mode of convergence of distributions) under a regularity condition. Second — the source paper of this candidate, 'On asymptotic normality in estimation after a group sequential trial' (Sequential Analysis, 2020) — they proved a much sharper theorem in the following model: outcomes are i.i.d. $N(\mu,\sigma^2)$ ; interim analyses of the partial sum occur at $k_1 n, k_2 n, \ldots, k_L n$ , with deterministic maximal trial length $k_{L+1} n$ , and with $L$ (the number of possible early stops) and the information fractions $k_i/k_{L+1}$ fixed while $n \to \infty$ ; at look $i$ , given that the trial has not stopped earlier, the conditional probability of stopping is $\psi(K_{k_i n}/(k_i n)^\gamma)$ , where $K_m$ is the partial sum, $\psi$ is a fixed measurable function, and the shape parameter $\gamma$ indexes the boundary family ( $\gamma = 1/2$ gives the Pocock paradigm of constant boundaries on the standardized scale, $\gamma = 0$ the O'Brien-Fleming paradigm of constant boundaries on the sum scale, and $0 \le \gamma \le 1/2$ the Wang-Tsiatis designs interpolating between them); the unconditional probability of stopping at look $i$ therefore carries the survival product $\prod_{j<i}(1-\psi(K_{k_j n}/(k_j n)^\gamma))$ . Their Theorem 6 states that $\sqrt{N_n}(\hat\mu_{N_n} - \mu)/\sigma$ converges in total variation to the standard normal law in each of five regimes (A)–(E), covering most combinations of $\gamma$ , $\mu$ , and regularity of $\psi$ , and Corollary 9 deduces that naive confidence intervals $\hat\mu_{N_n} \pm x\sigma/\sqrt{N_n}$ have asymptotically correct coverage for two-sided power boundaries $\psi(x) = \mathbf{1}\{|x| \geq C\}$ whenever $\mu \notin \{-C, C\}$ and $(\gamma, \mu) \neq (1/2, 0)$ .

Two features of the source paper delimit exactly what is missing. First, Remark 8 explicitly leaves the case $\gamma = 1/2$ , $\mu = 0$ — the Pocock boundary evaluated at the null hypothesis, precisely the parameter value at which type-I error is controlled — without any normality result, and Example 10 proves that normality genuinely fails there: the Kolmogorov distance (the supremum over all $x$ of the difference between the two cumulative distribution functions) between $\sqrt{N_n}(\hat\mu_{N_n}-\mu)$ and the standard normal stays bounded away from zero for all $n$ . In that regime the stopping rule remains genuinely random in the limit (the early-stopping probability converges to a nondegenerate value), and the limiting law is a computable non-normal mixture of truncated normals rather than $N(0,1)$ ; the paper's own calculation shows the law is exactly invariant in $n$ for a single interim look. Second, Example 11 shows the regularity conditions on $\psi$ cannot be omitted (for the discontinuous rule 'stop when the partial sum is negative' the Kolmogorov distance equals $1/8$ for every $n$ ). Related work underscores that the stopping effect need not wash out in other asymptotic regimes: Tarima and Flournoy (arXiv:1908.01411) show that under local alternatives with dense monitoring, test statistics after group sequential stopping converge to non-normal mixtures of truncated normal distributions; Grayling and Wason (Statistical Methods in Medical Research, 2023) compare nine bias-adjusted point estimators for two-stage GSTs at finite samples; and Aronowitz and Bartroff (arXiv:2606.07499, 2026) derive asymptotic multivariate normality and explicit non-asymptotic bounds, obtained via Stein's method (a technique for bounding the distance of a distribution to the normal law), for the joint law of the vector of quasi-maximum likelihood estimators (maximizers of a likelihood that may differ from the true data-generating model) evaluated at the deterministic interim times, under model misspecification and within-group dependence — a different object from the estimator at the random stopped time. None of these completes the characterization for post-trial estimation.

Problem Statement

Determine when estimation after a group sequential trial retains the asymptotic behavior it would have with a deterministic sample size — and what replaces that behavior when it does not. Concretely, in the regime where the maximal number of observations grows while the number of interim analyses and the information fractions are held fixed (as in Berckmoes–Ivanova–Molenberghs, Sequential Analysis, 2020), and within the source model — i.i.d. normal outcomes, stopping rules driven by the partial sums through a boundary function ψ and shape parameter γ — give necessary and sufficient conditions on (γ, µ, ψ) for the randomly stopped sample mean to be asymptotically normal, identify the limiting distribution wherever normality fails (e.g., the non-normal mixture limits at the Pocock-type boundary under the null), and thereby delimit exactly where naive confidence intervals and tests based on the post-trial estimator are asymptotically valid. This is precisely the gap the source paper leaves: its Theorem 6 supplies only sufficient regimes, its Examples 10–11 show that the conditions cannot be dropped, and necessity — together with the limit laws in the failure region — remains open. The source does not address extensions of this characterization beyond its own model; work on non-normal outcomes, stopping statistics other than the raw partial sum (standardized Z-statistic, score- or MLE-based boundaries, multi-parameter and multi-armed designs), a number of interim analyses that grows with the maximal sample size, unknown nuisance parameters such as σ, or estimators other than the sample mean (maximum likelihood, bias-adjusted, and median-unbiased estimators) constitutes natural further directions in the same spirit and is noted here as background, not as part of the open question left by the source.

References

Berckmoes, B., Ivanova, A., Molenberghs, G. (2020). On asymptotic normality in estimation after a group sequential trial. Sequential Analysis 39(4):443–466. DOI: 10.1080/07474946.2020.1826784; https://arxiv.org/abs/1805.03825 (arXiv:1805.03825)
Berckmoes, B., Ivanova, A., Molenberghs, G. (2018). On the sample mean after a group sequential trial. Computational Statistics & Data Analysis 125:104–118. DOI: 10.1016/j.csda.2018.03.016; https://arxiv.org/abs/1706.01291 (arXiv:1706.01291)
Berckmoes, B. (2018). An approach theoretic version of Anscombe's theorem with an application in biostatistics. Applied Categorical Structures 26(5):883–889. DOI: 10.1007/s10485-018-9512-8; https://arxiv.org/abs/1612.05104 (arXiv:1612.05104)
Berckmoes, B., Ivanova, A., Molenberghs, G. (2018). Conditional bias reduction can be dangerous: a key example from sequential analysis. arXiv:1812.06046; https://arxiv.org/abs/1812.06046
Milanzi, E., Molenberghs, G., Alonso, A., Kenward, M.G., Tsiatis, A.A., Davidian, M., Verbeke, G. (2015). Estimation After a Group Sequential Trial. Statistics in Biosciences 7(2):187–205. DOI: 10.1007/s12561-014-9112-6; https://doi.org/10.1007/s12561-014-9112-6
Tarima, S., Flournoy, N. (2019). Effect of Interim Adaptations in Group Sequential Designs. arXiv:1908.01411; https://arxiv.org/abs/1908.01411
Grayling, M.J., Wason, J.M.S. (2023). Point estimation following a two-stage group sequential trial. Statistical Methods in Medical Research 32(2):287–304. DOI: 10.1177/09622802221137745; https://doi.org/10.1177/09622802221137745
Aronowitz, J., Bartroff, J. (2026). Non-asymptotic bounds for quasi-MLE, misspecified models, and dependence under group sequential sampling. arXiv:2606.07499; https://arxiv.org/abs/2606.07499