GST hex sticker. A small forest plot with the global test marked in red.

Totality of Evidence in Minutes

Interactive Apps for GST Power and
Publication-Ready Forest Plots

Xiao Ni · R/Pharma 2026

Originally developed at Sarepta Therapeutics · Maintained by Xiao Ni

xni7.github.io/gst

The two headaches

The design headache

Several endpoints, correlated, with different effect sizes. Power depends on all of them — and on the test.

The reporting headache

Every change to a header or an interval colour means another round of plotting code. 

p <- forest_data() |>
  forest_plot(
    ci_color = "#008A99",
    display_cols = c(
      "Group", "forest",
      "est+ci", "p"))
# edit, re-run, re-export

Thumbnail of a multi-endpoint forest plot.

The two apps

Plan it. Power it. Publish it.
GST Power App (Plan it · Power it)
  • Size studies analytically from effects and correlations.
  • See when combining endpoints helps or hurts.
  • No repeated-trial simulation script for each scenario.
Forest Plot App (Publish it)
  • Flexible journal-style formatting.
  • No patient-level data (summary statistics only).
  • Export figures, data and R code to rebuild the plot.

Model-based analytical calculations support exploration. Simulation can still assess calibration.

Methods in {gst}

What is a GST?
A global statistical test assesses a joint signal across endpoints.Note: a significant global test does not mean every endpoint improved.
The weights
Equal-weight OLS or inverse-correlation GLS (O’Brien, 1984).
The cutoff
Degrees-of-freedom approximations: Dallow, O’Brien, Logan–Tamhane.
O’Brien Plus (GST_Any)
Power for the joint test and at least one individual endpoint being significant.

Also in {gst}: Wei–Lachin; gst::obt(), an O’Brien-based truncated closed-testing routine (Luo et al., 2025) — not in today’s demos. DOI: 10.1002/pst.70010

{gst} package structure

gst/ v0.3.1 · one R package behind both apps
├── R/ Package functions: obrien(), power & sample size, forest_plot(), forest_reprex(), Shiny modules
├── apps/ gst-power/ and forestplot/: deployment entry points (app.R + manifest.json)
├── inst/www/ App assets: CSS and logo
├── man/ Function reference (.Rd)
├── vignettes/ obrien · gst_power_app · forestplot · reference
├── tests/ testthat unit tests
├── pkgdown/ Documentation-site assets
├── DESCRIPTION Package metadata and dependencies
├── NAMESPACE Exported functions
└── app.R Local launcher: gst_power_app()

Docs: https://xni7.github.io/gst/ · Code: https://github.com/xni7/gst

Three designs, different power

Design Effects (δ₁, δ₂) Correlation ρ OES GST power Best individual
Balanced (0.4, 0.4) 0.3 0.50 94% 80%
Redundant (0.4, 0.4) 0.9 0.41 82% 80%
Uneven (0.5, 0.1) 0.8 0.32 60% 94%

Two endpoints, n = 100 per arm, two-sided α = .05, O’Brien OLS, Dallow df = 198. OES from gst::get_oes(); GST power is two-sample t power at the OES. Rounded as in the app; bold marks the higher power in each row. Best individual is a planning benchmark, not a multiplicity-adjusted selection rule.

Operational effect size (OES)

OES = Σδ √(ΣΣρij) Sum of individual effects Entries of the correlation matrix
Operational effect size Largest single effect

For equal-weight OLS, at the same n and degrees of freedom, the global test beats the best single endpoint when OES exceeds the largest single effect.

O’Brien OLS operational effect size, gst::get_oes(): 1ᵀδ / √(1ᵀR1). The three designs: 0.50 vs 0.40, 0.41 vs 0.40, 0.32 vs 0.50. A planning comparison, not permission to pick the test after seeing the data.

Sizing the study

Single endpoint
Effect 0.4 → about 100 per arm for 80% power.
GST, balanced design
Effects (0.4, 0.4), ρ = 0.3 → 65 per arm for 80% power.

✓ Calculated analytically with noncentral t distributions — no simulation grid needed to explore.

Sample-size widget from the gst power-app vignette. Effects 0.5 and 0.5, correlation 0.5, target power 0.8, Dallow, OLS. The curve crosses about 49 per arm.

Widget screenshot from the gst vignette: a different example — effects (0.5, 0.5), ρ = 0.5 → 49 per arm.

65 per arm from gst::get_n() in gst 0.3.1 (power 80.2%); single endpoint 80.4% at n = 100 (pwr::pwr.t2n.test). Two-sided α = .05, Dallow df.

Combining can help or dilute

Stouffer p-value plane. Cases 1 and 2 share one point at p = .08 and .08. Case 3 is a red triangle near the top left. Boundaries for correlation 0.3 (green, solid), 0.9 (amber, dashed), and 0.8 (red, dotted); the case 1 gain region is shaded.

Correlation-adjusted Stouffer
Case (p₁, p₂) ρ Combined p
1 · Helps (.08, .08) .3 .0261
2 · Misses .05 (.08, .08) .9 .0627
3 · Dilutes best (.00051, .48) .8 .0259
1 → 2: the point did not move; the .05 boundary did. Case 2 is still below .08.
3: below .05, but weaker than its best single p (.00051).

Hypothetical p-values; separate from O’Brien power. gst’s Stouffer: one-sided z, Nyholt effective-test scaling (gst:::p_combine). Curves: combined p = .05 boundaries. Shaded: case 1 gain. Axes use sqrt(p) spacing.

The forest plot

Input

Summary statistics (estimate, CI, p) via table, paste or CSV template. No patient-level data.

Forest plot of three endpoints with overall, 4–5 year, and 6–7 year rows. Only the starred 6–7 year row of endpoint 3 is red (inverted).

Formatting

Edit CI colours, column widths and headers interactively.

Transformations

Invert reverses the plotted direction; printed numbers stay as entered.

Standardize shows standard-error units; printed estimates and nominal p-values unchanged.

Built-in demo values from {gst}, not clinical results. Starred row: plotted direction reversed; printed estimate unchanged. Review against target-journal requirements.

From the app to R

Interactive edits (the app)

In the app

R Code Save Code Copy

Exports

CSV / RDS for data · PNG / JPEG for figures

Reproducible R (the script) 
library(gst)
p <- forest_data() |>
  forest_plot(
    ci_color = "#008A99",
    ci_color_inverted = "#C44B49",
    display_cols = c(
      "Group", "forest", "est+ci", "p"))
writeLines(forest_reprex(p),
           "forest-reproduce.R")

Compact workflow shown. The companion export rpharma-forest-example.R (gst 0.3.1) reproduces forest-plot.png; slide 9 shows a display copy of it.

Acknowledgements

Package authors and contributors
Lerong Li
Anthony Pileggi
Xiaolong Luo
Stephen Knapp
Preston Burns
Sarepta Therapeutics
Original development and support
Open-source community
R package maintainers and R/Pharma organizers

Plan it. Power it. Publish it.

{gst} takes you from trial design to the forest plot — size your GST study with the Power App, report it with the Forest Plot App.

GST Power AppSize the study Forest Plot AppReport the figure

Totality of evidence, in minutes.

Apps: xni7-gstpower.share.connect.posit.cloud · xni7-gstmakeforestplot.share.connect.posit.cloud
Source: github.com/xni7/gst · xni7.github.io/gst · Slides: xni7.github.io/talks/rpharma2026-gst/ · Video: Talk video (9.5 min)