🧫  Risk of Peeking

Risk of Peeking#

The inflated false-positive risk from repeatedly checking a fixed-horizon test early.

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 Despite careful review, it may still contain mistakes, omissions, or out-of-date information.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#

Peeking is looking at an experiment’s results before it officially ends and acting on that interim data. The risk of peeking is the inflated false-positive (Type I) rate that results from repeatedly checking and stopping as soon as significance appears.

Why it breaks the test#

A fixed-horizon test assumes a single look at a predetermined sample size. Each extra peek is another independent chance for noise to cross \(p < 0.05\), so the true error rate compounds far above the nominal \(\alpha\). Peek ten times at a 5% threshold and the real false-positive rate can reach 20–30% — this compounding is called alpha inflation.

What it costs#

Concretely: a button test shows A ahead on day 1 (\(p = 0.04\)), you stop and crown A — but over the full two weeks B would have won. The early stop produced a false conclusion, and at scale that means wrong launches, lost revenue, and eroded trust in experimentation.

How to avoid it#

Four routes: predefine the sample size and duration and only check at the end; use sequential testing / alpha-spending designs built for interim looks (group sequential, O’Brien–Fleming, Pocock); use Bayesian methods designed for continuous monitoring; or, if peeks are unavoidable, apply multiplicity corrections (Bonferroni, Holm). The peeking problem is precisely why the whole machinery of fixed-horizon and sequential testing exists.


Theme: A/B Testing & Experimentation  ·  All terminology



See also

Source article Adapted (context, re-expressed) in our own words from: Risk of Peeking (insightful-data-lab.com).

Tags: purpose: reference topic: terminology level: intermediate