Sequential Alpha Spending Planner

Lay out two-sided sequential testing boundaries for each interim look with either O'Brien-Fleming or Pocock alpha-spending so your experiment stops responsibly without inflating type I error.

Two-sided type I error you want across the whole experiment (%).
Total interim analyses including the final look.
Optional. Defaults to informational only—used to show sample checkpoints per variant.
Optional. Defaults to 0 for classic O'Brien-Fleming boundaries.

Validate sequential testing plans with your experimentation platform—implementation details (test statistic, sidedness) can change the exact boundaries.

Examples

  • Alpha 5%, 5 looks, max users 50,000 ⇒ Look 1: z ≥ 3.471 • alpha 0.0005 • reach ≈ 20.00% of 50,000 users • cum 0.0005 | Look 3: z ≥ 2.454 • alpha 0.0141 • reach ≈ 60.00% of 50,000 users • cum 0.0174
  • Alpha 5%, 4 looks, Pocock style ⇒ Look 1: z ≥ 2.413 • alpha 0.0158 • cum 0.0158 | Look 4: z ≥ 2.413 • alpha 0.0158 • cum 0.0632

FAQ

Are the boundaries two-sided?

Yes. The calculator assumes two-sided tests. For one-sided boundaries divide the returned alpha values by two before configuring your platform.

How should I pick the number of looks?

Use the cadence that matches your product release cycle. More looks add flexibility but require larger z-thresholds early on.

Can I adjust information fractions?

This version assumes equally spaced information. If your sample ramp is uneven, scale the reach percentages manually or rerun the math with custom info fractions.

How should I combine this with group sequential sample size calculations?

First size your trial using the same number of looks and alpha spend, then plug the maximum per-variant sample into this tool so checkpoint percentages align with the underlying design.

Additional Information

  • O'Brien-Fleming keeps early stopping boundaries very conservative and releases most alpha near the final look.
  • Pocock uses a constant z-threshold at every look, which increases the chance of early stopping but inflates midstream alpha usage.
  • Cumulative alpha values are approximate because spending functions are continuous—use them as guardrails when configuring experiment platforms.
  • Result units: z-thresholds and alpha spent per look (two-sided)

Related calculators