🧫  Fixed-Horizon Testing

Fixed-Horizon Testing#

Testing where the sample size is fixed in advance and analysed only at the end.

Important

✨ AI-generated content. This page was written with the assistance of an AI language model and is provided as a learning aid. Despite careful review, it may still contain mistakes, omissions, or out-of-date information. Whether you are new to the topic, a team lead, or a senior practitioner, treat it as a starting point rather than an authoritative reference: read it critically and independently verify anything you act on (code, commands, figures, and factual claims) against official documentation and primary sources before relying on it.

What it is#

Fixed-horizon testing is the general statistical principle behind the classic experiment: decide the sample size \(n\) (or end time) in advance, collect data to that point, and run the hypothesis test exactly once. The “horizon” is that pre-committed stopping point. Applied to an A/B test specifically, this is the traditional A/B test.

Why the horizon matters#

Fixing it up front is what keeps the statistics honest: the Type I error stays at \(\alpha\), peeking is ruled out (checking early and stopping on significance inflates false positives), and the resulting p-values and confidence intervals remain valid under their assumptions.

The procedure#

State \(H_0\) and \(H_1\); choose \(\alpha\) (say 0.05); run an a-priori power analysis to size the sample; fix the horizon; collect to it; test once; decide.

Example#

To detect a +10% lift (5% → 5.5%) at \(\alpha = 0.05\) and power 0.80, an a-priori power analysis calls for ≈ 7,850 users per variant. You stop at that horizon and run a single two-proportion z-test: \(p \le 0.05\) rejects \(H_0\), otherwise you fail to reject.

Strengths and limits#

It is simple, widely accepted, and preserves error guarantees. The price is rigidity: no early stop even when the result is already obvious, wasted samples when an effect is large, and no continuous monitoring. Sequential and adaptive methods (α-spending, Bayesian updating, bandits) trade some of that simplicity for the ability to stop early.


Theme: A/B Testing & Experimentation  ·  All terminology



See also

Source article Adapted (context, re-expressed) in our own words from: Fixed-Horizon Testing (insightful-data-lab.com).

Tags: purpose: reference topic: terminology level: intermediate