exact-binomial-spending

name: exact-binomial-spending description: Build and document exact binomial group sequential designs with spending in gsDesign. Use when the user mentions toBinomialExact, gsBinomialExact, repeatedPValueBinomialExact, sequentialPValueBinomialExact, simBinomialSeasonalExact, vaccine/prevention efficacy with seasonal looks, or blinded information-based adaptation with fixed spending fractions.

Exact Binomial Spending Workflow

Use this skill for exact-binomial-with-spending design, analysis, testing, and vignette authoring in this repository.

When To Apply

Apply this skill when the task includes one or more of:

• toBinomialExact(), gsBinomialExact(), or exact event-count bounds.
• repeatedPValueBinomialExact() or sequentialPValueBinomialExact().
• simBinomialSeasonalExact() for seasonal fixed/adaptive simulation summaries.
• Spending functions for exact binomial monitoring.
• Seasonal vaccine/prevention workflows with analyses by season.
• Blinded information-adaptive enrollment updates with fixed spending fractions.
• New vignette authoring for exact binomial spending scenarios.

Repository Conventions To Preserve

1. Spending denominator is planned final events

• Compute spending time relative to planned final event count.
• Do not change the denominator when observed events exceed plan.
• If explicit overrides are needed, use usTime/lsTime consistent with gsDesign()/gsSurv() conventions.

2. Single look at or beyond final planned events

• Enforce at most one analysis with n.I >= planned_final_events.
• Rationale: spending functions already saturate at the first t >= 1.

3. Exact efficacy control under non-binding futility

• For Type I calculations of efficacy bounds, ignore non-binding futility.
• If reporting futility behavior from simulation, treat futility stopping probability separately from non-binding Type I error.

4. Public vs internal API

• Keep user-facing functions exported and documented.
• Keep low-level bound helpers internal (@keywords internal, @noRd) unless explicitly requested.

5. Official update path at analysis time

• Use toBinomialExact(gsD, observedEvents = ...) for analysis-time exact bound updates.
• For explicit spending-time control, use toBinomialExact(gsD, observedEvents = ..., usTime = ..., lsTime = ...).
• For selective futility designs, preserve testLower intent from gsSurv().
• If simulation needs bound checks at fixed alpha, prefer public wrappers/helpers over gsDesign::: in user-facing examples whenever practical.

Standard Implementation Pattern

1. Build time-to-event scaffold with gsSurv(...) and chosen spending function.
2. Convert timing/counts to integer events (toInteger(...) as needed).
3. Convert to exact binomial bounds with toBinomialExact(...).
4. Compute exact p-values with:
   • repeatedPValueBinomialExact(...)
   • sequentialPValueBinomialExact(...)
5. For seasonal simulation studies, prefer simBinomialSeasonalExact(...) instead of embedding large custom simulation helpers directly in vignettes.
6. Add focused tests for:
   • input validation,
   • spending-time constraints,
   • coherence (sequential == min(repeated)),
   • behavior under planned and updated event counts.
7. Regenerate docs and namespace with roxygen.

Vignette Authoring Pattern

For exact-binomial spending vignettes, include:

1. Assumptions section

• Clinical context, null/alternative effect scale, event/dropout assumptions.

2. Design section

• Planned seasonal looks and spending times.
• Exact boundary derivation and spending checks.

3. Inference section

• Demonstrate repeated and sequential exact p-values on plausible observed counts.
• Include a short demo of toBinomialExact(observedEvents = ...) for analysis-time bound updates.

4. Adaptive extension

• Blinded update rule for future enrollment/information.
• Keep spending fractions unchanged.
• If final-look full spending is desired under under-run, use maxSpend = TRUE (or final_full_spending = TRUE in simulation helper).

5. Simulation section

• Prefer package simulation helpers (for example, simBinomialSeasonalExact()).
• Lightweight runnable chunk for package builds.
• eval=FALSE offline template for larger runs.
• Present both efficacy crossing probability and futility stopping probability; when quoting Type I error under non-binding futility, label that explicitly.

6. Operational note

• Defer NEWS.md updates until feature batch is stable, unless the user asks otherwise.

Prompt Templates

Use ready-to-copy prompts from:

• examples.md