[xidongdxi]

Adds vignettes/group-sequential-testing-rpact.Rmd — "Graphical Approaches in Group Sequential Design: Combining graphicalMCP and rpact" — the rpact counterpart of the gsDesign vignette, mirroring it section for section. Targets the rpact-engine branch, where graph_test_shortcut_rpact() lives.

Contents

Reproduces both case studies of the group sequential testing vignette with graph_test_shortcut_rpact() and shows how rpact is used inside the function:

• Spending and boundaries from getDesignGroupSequential() (alphaSpent, criticalValues, stageLevels), with the mapping of graphicalMCP's spending functions to rpact's design types.
• Repeated p-values as the root of c_k(alpha) = Z_k on the design's critical values, shown with uniroot() and compared with rpact::getRepeatedPValues() (agree to 8.5e-7); sequential p-values as the cumulative minimum.
• Diabetes case study (Maurer & Bretz 2013, 2 of 3 analyses, look_back = FALSE), including why a schedule may end before full information for alpha spending designs but not for the classical boundary families, the key inequality, test_values, and the boundary table.
• Oncology case study (1–3 analyses per hypothesis, look_back = TRUE), per-hypothesis look_back, and the "when look_back makes a difference" scenario.
• Spending time via the two-step "asUser" construction, monitoring under changed final information with the three-scenario boundary comparison done in pure rpact, and the equivalence of over-running information fractions.
• Additional examples: design types per hypothesis, rpact's alpha spending types on the diabetes trial, the classical families ("OF", "P", "WT") on the oncology trial (whose schedules run to 1), and user-defined spending via "asUser" — including an explicit demonstration of its limit for spending functions whose shape depends on the level, with a comparison against graph_test_shortcut_gsd().

Also registers the vignette in _pkgdown.yml and NEWS.md, and adds rpact's, graphicalMCP's, Cham to inst/WORDLIST.

Verification

• Renders without warnings or errors (rmarkdown::render under pkgload::load_all).
• Every example was cross-checked against the mvtnorm engine (graph_test_shortcut_gsd()): identical rejections and decision analyses, repeated/adjusted p-values agreeing to about 1e-6, boundaries to 1e-5 or better — except the "asUser" piecewise example, where the difference is the point being made.
• devtools::spell_check(): the only words flagged in the new vignette are ge/le from LaTeX \ge/\le, which are flagged identically in the existing gsDesign and main vignettes.

Wording of the classical-family requirement

The classical boundary families need the final information fraction to be exactly 1 because their boundary constant is calibrated over the whole planned schedule — a property of the designs, not an rpact rule (graphicalMCP's own spending_wt() imposes the same requirement). The second commit rewords the engine's error message, its documentation, and the vignette to say so.

Note

Writing the vignette surfaced a pre-existing bug unrelated to rpact: test_input_val() mishandles single-hypothesis graphs (weighting_strategy[, 1] drops to a vector, so nrow() is NULL) — warnings in graph_test_shortcut()/graph_test_closure(), a hard error in graph_test_shortcut_gsd(). The vignette avoids one-node graphs; the bug is tracked separately.

🤖 Generated with Claude Code