Composed Adaptive Pipeline — Type I Error Simulator

Estimates the trial-wide type I error rate when several adaptive mechanisms — historical borrowing, Bayesian sequential monitoring, sample-size re-estimation, and response-adaptive randomization — are combined in a single two-arm binary trial over overlapping interim analyses. It illustrates that acceptable component-level operating characteristics do not guarantee pipeline-level type I error control when the mechanisms are composed.

How it works

You enable any combination of the four mechanisms, set a null scenario (including optional prior-data drift and a linear time trend), and the tool runs a Monte Carlo simulation to estimate the pipeline-level rejection rate under the null, with a Monte Carlo standard error and confidence interval. It reports how far the simulated rate sits from the nominal level and the super-additive interaction between mechanisms.

When to use it

  • You are combining multiple adaptive features in one trial and want to check the composed type I error rather than assume component-level control carries over.
  • You want to stress-test a design under prior-data conflict and temporal drift.
  • You want to reproduce or explore the composition finding for your own configuration.

Assumptions & limitations

  • This is a simulation study, not a general theorem: the reported inflation (e.g. a roughly threefold increase in the headline scenario) arises under a specific configuration — mild prior-data conflict plus a linear time trend with all four mechanisms active — and it grows or shrinks with the conflict, trend, and settings you choose.
  • The central point is that acceptable component-level operating characteristics do not guarantee pipeline-level type I error control under composition; it does not claim composition always inflates error, and a no-conflict, no-trend configuration remains approximately controlled.
  • The model is a two-arm binary trial with specific mechanism variants (blinded pooled SSR, Thompson-sampling RAR, a MAP-mixture prior, a linear trend) — other implementations may behave differently.
  • Results are Monte Carlo estimates for the scenarios you simulate, not guarantees for a realized trial.

For the full methodology, derivation, and worked examples, see the guide for your design:

Loading...