Why This New Statistics in Medicine Paper Matters for Adaptive Multi-Arm Trials

A new open-access paper by Silke Jörgens and Gernot Wassmer highlights key trade-offs in adaptive multi-arm two-stage designs—and shows why rpact is such a powerful framework for exploring them.

rpact
conditional dunnett
inverse normal design
adaptive designs
multi-arm trials
statistics in medicine
Author

RPACT Team

Published

March 16, 2026

A new open-access article by Silke Jörgens and Gernot Wassmer has just been published in Statistics in Medicine:

Jörgens S, Wassmer G (2026). A Note on “A Comparison of Two Methods for Adaptive Multi-Arm Two-Stage Design”. Statistics in Medicine, 45:e70448. DOI: 10.1002/sim.70448

For us at RPACT, this publication is especially meaningful for two reasons.

First, Gernot Wassmer is not only a long-standing leader in adaptive clinical trial methodology, but also a core part of the RPACT team. Second, the paper explicitly points readers to rpact as the software framework used for the simulation work and as a practical route for reproducing and extending these kinds of analyses.

This is exactly the kind of visibility we like to see: strong statistical methodology, transparent discussion of relevant design choices, and open-source software that makes these methods accessible in practice.

What is the article about?

In this note, the authors comment on a recent comparison of two approaches for adaptive multi-arm two-stage designs with many-to-one comparisons:

  • the inverse normal combination test approach, and
  • the conditional Dunnett approach.

Both methods are well-established and both can be embedded in a closed testing framework for confirmatory adaptive designs. The article does not argue that one method is universally superior. Instead, it sharpens the discussion around where observed power differences actually come from.

That matters, because in adaptive dose-finding or treatment-selection settings, it is often tempting to focus only on headline power values. But for real design work, the deeper question is usually:

Why do two valid procedures behave differently, and when does that difference really matter in practice?

That is precisely the question addressed here.

The key message in plain language

One of the central takeaways of the article is that the often-reported power advantage of the conditional Dunnett approach over the inverse normal approach should be interpreted with more nuance.

The authors show that:

  • the conditional Dunnett approach can show a power advantage,
  • but that advantage is often small or negligible in practically relevant settings, especially when only the best treatment arm is selected, and
  • the main reason for the difference is not primarily non-consonance, but rather loss of power in the global intersection test when p-values are calculated stagewise rather than from the full cumulative information.

This is an important clarification.

In other words, the discussion is not just about “which method wins.” It is about understanding the mechanics of adaptive multiple testing procedures and recognizing that the practical impact depends strongly on the selection strategy and the structure of the design.

Why that matters for trial design teams

For statisticians working on adaptive platform, dose-response, or treatment-selection trials, this is highly relevant.

In practice, teams need methods that are not only theoretically sound, but also:

  • transparent,
  • reproducible,
  • flexible enough for realistic adaptations, and
  • implementable in a way that supports simulation-based design decisions.

The paper highlights a point that practitioners often appreciate immediately: if only one treatment arm is selected at interim, the power difference between the two approaches is negligible in the investigated settings. Even when selecting the best two arms, the difference is often still small.

That means the “best” approach cannot be chosen responsibly from a simplified headline such as “method A has more power than method B.” The design context matters.

Why the inverse normal framework remains so attractive

A particularly valuable aspect of the article is that it does not stop at reporting power differences. It also reminds readers why the inverse normal combination test framework remains highly attractive.

The article highlights several advantages:

1. Natural extension beyond two stages

The inverse normal method extends naturally to more than two stages. That is highly relevant in modern adaptive development programs, where multi-stage decision-making is often part of the strategic design space.

2. Greater flexibility when variance is unknown

The conditional Dunnett approach is restricted to the case of known variance, which can be a strong assumption, especially in smaller studies. By contrast, the inverse normal framework naturally accommodates settings with unknown variance by using stagewise variance estimates.

3. A broadly accepted general framework for adaptation

The closed combination test is widely recognized as a powerful and flexible framework for deriving group sequential procedures that allow for general, data-driven adaptations.

For teams designing confirmatory adaptive studies, that flexibility is often not a side detail. It is the core reason these methods are attractive.

Where rpact comes in

This is where the publication becomes especially relevant for R users and for teams working in regulated or semi-regulated development environments.

The article explicitly states that the simulation study was performed in rpact, and it points readers to additional code examples and vignettes. The supporting information even includes a minimal rpact example illustrating how to work with:

Show the code of the minimal rpact example 
library(rpact)

set.seed(38448409)

assEffects <- c(0, 0, 0, 0.25, 0)

# note: group 5 is the control group
dataSet <- getDataset(
    means1 = rnorm(2, assEffects[1], c(1 / sqrt(100), 1 / sqrt(100))), 
    means2 = rnorm(2, assEffects[2], c(1 / sqrt(100), 1 / sqrt(100))), 
    means3 = rnorm(2, assEffects[3], c(1 / sqrt(100), 1 / sqrt(100))), 
    means4 = rnorm(2, assEffects[4], c(1 / sqrt(100), 1 / sqrt(100))), 
    means5 = rnorm(2, assEffects[5], c(1 / sqrt(100), 1 / sqrt(100))), 
    stDevs1 = c(1, 1),
    stDevs2 = c(1, 1),
    stDevs3 = c(1, 1),
    stDevs4 = c(1, 1),
    stDevs5 = c(1, 1),
    n1 = c(100, 100),
    n2 = c(100, 100),
    n3 = c(100, 100),
    n4 = c(100, 100),
    n5 = c(100, 100)
)

getDesignConditionalDunnett(
    secondStageConditioning = TRUE
) |>
    getStageResults(
        dataInput = dataSet,
        intersectionTest = "Dunnett",
        normalApproximation = TRUE
    ) |>
    getClosedConditionalDunnettTestResults()

getDesignInverseNormal(
    kMax = 2,
    typeOfDesign = "noEarlyEfficacy"
) |>
    getStageResults(
        dataInput = dataSet,
        intersectionTest = "Dunnett",
        normalApproximation = TRUE
    ) |>
    getClosedCombinationTestResults()

That is a strong signal. The methodology is not only discussed at a theoretical level—it is also connected to a practical, openly available implementation.

For readers who are new to rpact, this is worth emphasizing:

rpact is not just a tool for standard group sequential calculations. It is a serious framework for confirmatory adaptive design and analysis.

It allows users to move from conceptual design discussions to executable simulations and reproducible operating characteristics with remarkably little friction.

What this says about rpact

Publications like this help illustrate what makes rpact powerful.

rpact connects theory and implementation

Many papers discuss adaptive methods in abstract notation only. rpact helps turn those ideas into concrete design objects, simulation workflows, and interpretable outputs.

rpact supports advanced adaptive methodology

This paper is about nontrivial methodology: multi-arm, two-stage, adaptive selection, closed testing, Dunnett-based procedures, conditional error logic, and inverse normal combination testing. The fact that rpact is used directly in that context tells you a lot about the package’s scope.

rpact makes comparison possible

A major strength of rpact is that it enables users to compare design options in a disciplined way. That is essential when the right answer is not “always use method X,” but rather “understand the trade-offs for your trial.”

rpact is open and visible

The paper is open access, and rpact is open source. That combination matters. It supports transparency, reproducibility, and broader methodological uptake across academia, biotech, pharma, and consulting.

Our perspective at RPACT

We believe this article is a nice example of how modern statistical software should support methodological work.

Good software should not hide complexity behind black boxes. It should make advanced methods usable, explainable, and reproducible. That has been one of the guiding ideas behind rpact from the beginning.

This publication also reinforces something we see repeatedly in practice: sophisticated adaptive methodology becomes much more valuable when teams can actually simulate it, stress-test it, and communicate it clearly. That is where a well-designed software framework makes a real difference.

Read the article and explore rpact

If you work on adaptive clinical trial design, this paper is well worth reading. It is concise, technically sharp, and practically relevant.

And if the topic sparks ideas for your own work, rpact gives you a direct way to explore them yourself.

You can start here:

At RPACT, we are proud to see rpact contributing to published methodological work of this caliber—and we hope this encourages even more teams to look at what is possible with modern open-source software for confirmatory adaptive designs.