Composed Pipeline T1E
PROBETAPipeline-level T1E for composed adaptive designs (MAP + monitoring + SSR + RAR) — JSM 2026 reproduction.
Adaptive Mechanisms
Data-Generating Scenario
Empty = H0 (rate equals pc)
γ* is calibrated at the anchor p0 = 0.25 (not the mean of any prior in play — the mixture mean is 0.3132). Raising p is a departure from the calibration anchor. Trend δ shifts both arms in parallel, so it does not by itself break the null.
pt = pc, so this run is under H0 and its rejection rate is Type I error.
Design
The SSR rule uses the pooled response only to estimate the blinded nuisance level, holds Δ fixed, and assumes working 1:1 allocation. With the manuscript defaults it preserves the original N = 90 Wald planning power (0.5296) and maps the complete n = 45 pooled-count support to final N = 90–100.
Interim enrollment counts, ascending, each ≤ nbase. Re-scales automatically when nbase changes. The post-SSR final analysis is appended by the engine, so it is not listed here.
Control-arm MAP prior as a Beta mixture Σ wi · Beta(ai, bi). Defaults are the paper's RBesT fit to the historical cohorts. Weights must sum to 1. The treatment arm always uses Jeffreys — borrowing applies to control only.
Mixture mean ≈ 0.3132; moment-matched mixture ESS ≈ 5.5 (matched to one Beta).
Dominant component (w = 0.5887): mean ≈ 0.2568, ESS ≈ 71.7. A flat robustifying component pulls the mixture ESS well below this.
Weights sum = 1.0000
Reproduce JSM 2026 paper
Each preset runs a multi-scenario batch and tabulates the results against the paper's reported values. Single-seed T1E drifts by a few MCSE because this engine uses NumPy PCG64 where the paper used R's Mersenne Twister — raise N in the Advanced panel to tighten it. In the current five-seed, N = 5,000-per-cell verification, all 23 published cells are within three combined Monte Carlo standard errors.