A continuing purpose
gsDesign brings group sequential methods into an open, programmable environment for clinical trial design. Its central purpose is not just to calculate a set of stopping boundaries, but to help statisticians understand the resulting trial: its power, error spending, expected and maximum sample size, and the treatment effects and conditional power associated with its boundaries. That purpose connects the early package to its more recent survival, exact binomial, and monitoring tools.
This account follows selected developments rather than every release. It draws on Keaven Anderson’s recollections, the package’s recorded acknowledgments, release notes, and source history. It is also an invitation to add the recollections that those records cannot capture. The history is recorded through September 2026; features identified as development work should not be assumed to be in an installed CRAN release.
Foundations and the early package
The computational history predates the R package. Keaven wrote the original C code in 1999, based on the numerical integration routines described in Jennison and Turnbull’s Group Sequential Methods with Applications to Clinical Trials (Jennison and Turnbull 2000), published in 2000. Those routines supported multiple methodological papers before becoming the computational foundation of this initial gsDesign R package in 2006/2007. The package combines R interfaces with compiled routines for boundary crossing probabilities and boundary construction, so the same calculations can support both familiar designs and new ways of specifying them.
One example is Anderson’s Optimal Spending Functions for Asymmetric Group Sequential Designs (Anderson 2007). Published in 2007, with a manuscript received in 2005, it compared optimized spending function designs with designs not restricted to a spending function family. The paper minimized expected sample size, or its square, averaged over a prior distribution while retaining specified Type I and Type II error rates. This illustrates an early use of the numerical routines to go beyond deriving a standard boundary and instead ask which design best serves a stated objective.
An important organizing idea is error spending. Lan and DeMets (Lan and DeMets 1983) provided a way to specify cumulative error expenditure as information accumulates, rather than fixing a boundary only for a single predetermined analysis schedule. In gsDesign, alpha- and beta-spending help separate a monitoring policy from its boundaries at particular analyses. They allow the design to be evaluated at revised information times, subject to the assumptions of the sequential procedure. This is flexibility in implementation, not permission for unrestricted outcome-dependent changes to the design.
This separation also clarifies the idea of spending time. Information-based correlations are still used to define the joint multivariate normal calculations for boundary crossing, while the error spending rule itself need not be directly tied to statistical information. Lan and DeMets illustrated this distinction with calendar-based spending (Lan and DeMets 1983). The broader spending time concept has since been used in many trials and is the subject of a forthcoming publication.
A common application bases spending at each interim analysis on the minimum of planned and actual information. When minimum follow-up, accrual, or dropout means that an analysis requires more events than originally planned, this cap prevents the earlier analysis from spending alpha as though all of that extra information had been scheduled in advance. It retains adequate alpha-spending for the final analysis while still allowing the interim calculation to use the information actually observed.
Keaven Anderson has led the package’s statistical development and maintenance. The README credits Bill Constantine and Rich Calaway, then at Revolution Computing, with substantial early development, testing, and documentation help. These contributions established that documentation and checking were part of the software, not additions to it. Source: recorded acknowledgments.
The surviving repository contains gsDesign material from September 2008; for example, a manual was added on September 1. The CRAN source archive currently starts with version 2.0-5, dated July 1, 2009. Neither the earliest surviving archive file nor a repository timestamp alone establishes the first public release date. In particular, the repository’s 2007 initial commit is an R-Forge template, not evidence of a gsDesign release in that year.
Selected milestones
The version labels and descriptions below follow the package release
notes in NEWS.md. Release note dates describe the recorded
development history; they are not necessarily identical to the dates of
CRAN publication.
| Period / version | Development | Why it mattered |
|---|---|---|
| 1999 | Original C integration routines written by Anderson | A computational foundation for methodological research and the later R package. |
| 2007 | Optimal spending function paper, submitted in 2005 | Established numerical calculations used to compare and optimize asymmetric designs. |
| 2008–2009 | Early repository material and archived CRAN packages | A durable public record of the software and its documentation. |
| 2012 / 2.7 series | Introduction of nSurv() and
gsSurv()
|
Lachin-Foulkes survival sample size calculations connected to group sequential design, enrollment, dropout, and follow-up. |
| 2014 / 2.9-3 | Trimmed, gapped, and step spending tools | More control over when monitoring is active and how spending is specified. |
| 2017 / 3.0-4 | First GitHub-based release | A transition from the earlier R-Forge workflow. |
| 2019 / 3.1.0 | testthat, roxygen2, R Markdown, and pkgdown | Tests, help, examples, and a website became a more integrated development workflow. |
| 2020–2021 / 3.1.1–3.2.0 | Continuous integration and expanded tests | More systematic checking, with reported coverage above 80% in 3.2.0. |
| 2022–2023 / 3.3.0–3.6.0 | Vaccine efficacy examples, exact binomial conversion, integer conversion, and calendar-time survival designs | More direct connections between statistical designs and trial implementation. |
| 2026 / 3.9.0 | Schoenfeld and Freedman survival sample size options | Easier comparisons with other software while retaining the Lachin-Foulkes default. |
| 2026 / 3.10.0–3.11.0 and subsequent development | Survival scenario evaluation, selective monitoring, harm bounds, and target-based spending calibration | A broader set of design questions built on established computational tools. |
Making the software easier to maintain and use
The late 2018 and early 2019 modernization was enabled in large part
by Devin Pastoor, Harsh Baid, and Jonathan Sidi at Metrum Research
Group. Version 3.1.0 replaced RUnit testing with testthat, converted
help generation to roxygen2 and vignettes to R Markdown, and added the
pkgdown website. It also added sequentialPValue() and
extended spending time capabilities. The changes made worked examples,
testing, and reference documentation easier to maintain together.
Yilong Zhang contributed the GitHub continuous integration work for version 3.1.1. Collaboration with Cytel during 2020 expanded the test suite; the 3.2.0 release notes report coverage above 80%. Coverage is evidence about which code was exercised, not proof of statistical correctness. Independent calculations, regression tests, and examination of design properties remain important complements. Sources: acknowledgments and release notes.
Access has also expanded beyond writing R code directly. Anderson developed the original Shiny interface in extensive collaboration with John Lueders. Nan Xiao subsequently undertook the interface remake, and Yihui Xie has made ongoing improvements. This work supports interactive exploration, saving and reloading designs, and generating reproducible R code and R Markdown reports. The evolving interface is available as gsDesign Explorer. The link points to the maintained application rather than a frozen copy of the original interface.
The associated book, Group Sequential Design in R: A Practical Guide to the gsDesign Web Interface, by Anderson and Nan Xiao, connects interactive design exploration to the underlying methods. The complementary gsDesign Technical Manual provides a more detailed account of the package and its calculations.
The people named here are not an exhaustive list. Contributions include statistical suggestions, bug reports, testing, documentation, interface work, and maintenance as R and its dependencies change. The contributor history and individual issues and pull requests provide additional attribution; commit counts alone do not measure the value of those contributions.
From statistical information to a trial that can be run
Time-to-event sample size was a major extension of the package. The
nSurv() and gsSurv() routines were introduced
in the 2.7 series in 2012: the May
24, 2012 source addition includes their implementation and
documentation describing them as new in version 2.7.
gsSurv() connected the survival sample size calculations in
nSurv() to the group sequential calculations in
gsDesign(). This brought enrollment, dropout, calendar
time, follow-up, and stratification into an integrated survival design
workflow.
The Lachin-Foulkes framework (Lachin and
Foulkes 1986) remains the default for these sample size
calculations. Its implementation is a distinctive capability of
gsDesign, not generally offered by other survival design packages. This
made direct software comparisons difficult when the other software used
a different sample size approximation. A key 2026 addition, in
version 3.9.0, was therefore the Schoenfeld
option: users can select the same method in gsSurv() and
other software, making comparisons easier to interpret. The less used
Freedman method was enabled as well, alongside
Bernstein-Lagakos. Lachin-Foulkes remains the default; the alternatives
facilitate comparisons rather than replace it. The technical manual
illustrates comparisons
with rpact, which offers Schoenfeld and Freedman but not the
gsDesign Lachin-Foulkes method.
The 2023 releases added toInteger(),
toBinomialExact(), and gsSurvCalendar(). These
address different implementation questions: rounding sample sizes and
events, constructing exact event count boundaries for low event rate
studies, and scheduling survival analyses in calendar time. The exact
binomial method with spending was motivated by a user request and is now
available for vaccine and prevention studies. This is a separate
probability model with its own assumptions, not simply a cosmetic
relabeling of a normal boundary; exact boundary crossing calculations
retain the discreteness of the event counts.
In 2026, gsSurvPower() made evaluation of a specified
trial under alternative assumptions more explicit. Selective efficacy,
futility, and harm monitoring expanded the possible schedules.
Subsequent development refined exact conversion and follow-up
calculations, including the distinction between computing median
follow-up at a cutoff and finding a cutoff that achieves a target. These
developments share a practical theme: distinguish the original design
assumptions from what is held fixed when evaluating another scenario.
See the survival overview (vignette("SurvivalOverview"))
and enrollment planning examples
(vignette("SurvivalEnrollmentPlanning")).
A growing family of packages
The gsDesign2 package has been available for several years, with GitHub releases recorded in 2022. It extends trial design to settings where treatment effects may vary over time or across strata, including delayed benefit and other departures from proportional hazards. Its survival specifications allow stratified enrollment, event, and dropout assumptions, so non-constant effects can be considered when deriving and evaluating efficacy and futility bounds. See its release history and methodological vignettes.
The largely AI-developed gsDesignNB package adds designs for recurrent events with negative binomial outcomes to the family. It provides sample size, simulation, and interim monitoring tools for event rates, drawing on gsDesign for spending function and boundary calculations. Anderson describes its development as another example of using AI to build on established computational tools. The related packages extend the range of models and trial questions rather than requiring every endpoint to be handled by the same design function.
New questions, established calculations
Recent development allows a statistician to start with an interpretable interim target and solve for spending function parameters. Examples include conditional power, posterior predictive power, and approximate treatment effects at bounds. The result is still a spending function, with its associated flexibility when the information schedule changes. It must still be assessed as a complete design, including overall power, error probabilities, sample size, and effect sizes at the boundaries. A target achieved at the planned interim need not remain numerically identical at a different interim information fraction.
The conditional power spending vignette
(vignette("CPFutilitySpending")), Dragalin related methods
(vignette("DragalinFutilitySpending")), and effect-targeted
spending vignette (vignette("EffectSpending")) document
these development features. They illustrate the value of separating a
new calibration problem from the established probability calculations
used to evaluate its answer.
AI-assisted development has become part of this recent history. Anderson reports that much of the initial CP/PP spending calibration work was developed interactively with GPT-6 Astra High in Codex over a long weekend. This is an account of that development experience, not a measured estimate of productivity or a substitute for validation. The important enabling resource was the accumulated package: existing numerical routines, design objects, summaries, examples, and tests. New calibration code could be checked by reconstructing designs and evaluating them with existing functions. Such checks are especially valuable, although shared numerical components also make independent checks and code review important.
The numerical foundation itself has also been revisited. Nan Xiao used AI and Lean in research on the integration algorithms, combining numerical comparisons and performance measurements with Lean 4 / Mathlib proofs of properties of the guarded Newton update used in boundary finding.
The research led to internal optimizations and a faster, more
accurate Gauss-Legendre integration alternative in the benchmarked
designs. To preserve the established numerical behavior for existing
users, the new quadrature is opt-in, selected with
options(gsDesign.quadrature = "gl"); the default remains
"jt", the Jennison-Turnbull grid. Some internal speed
improvements have been adopted while retaining that default method.
Keeping the alternative optional protects continuity for a long-used
package without preventing users from benefiting from the new
algorithms. Details are recorded in the integration
change and the recursion
optimizations.
A visual record and an open history
A Gource video of the version control history was linked in the proposal for this vignette. It offers another way to visualize and explore repository activity. A future update could accompany this account; repository activity should not be read as a chronology of scientific ideas or a ranking of contributors.
This first account deliberately leaves room for the original
motivation, early collaborations, and practical experiences that are not
captured in release notes. Corrections and additions are welcome through
the issue tracker. For the current interface and complete changelog, use
the function reference and NEWS.md rather than this
selective history.