Statistical Power

A test needs enough people and enough time before its result means anything.

Test duration depends on traffic, baseline rate, and minimum detectable effect.

Statistical Power, side by side

Under-powered tests read noise as a winner. Size the sample first, and don't stop the moment it looks good.

How to apply it

Use a sample-size calculator before launching. Stopping early inflates false positives.

Statistical Power in practice: a concrete example

The test you stopped reading too early

An A/B test shows the variant up 12% after two days and it is tempting to ship it. But with low traffic and a small baseline rate, that early lead is mostly noise, and calling it now inflates the chance of a false win. Statistical power is about running long enough (traffic, baseline rate, and the effect you want to detect together) to tell a real difference from randomness. Put the numbers into a sample-size calculator before you launch, then wait for it.

Common Statistical Power mistakes

Peeking at the results and stopping the moment the line looks good, which sharply inflates false positives.
Running a test on traffic too thin to ever detect the effect size you care about, so it is inconclusive by design.
Ignoring the minimum detectable effect: a test powered to catch a 20% lift will call a real 3% lift a tie.

When Statistical Power doesn't apply

Formal power analysis is for decisions that hinge on a small, measurable difference. For a large, obvious change, or a product with too little traffic to ever reach significance, waiting months for a p-value is the wrong tool, and a qualitative read or a clear product judgment serves better. Rigour is worth it in proportion to the cost of being wrong.

Practise it on a real challenge

CTA Copy — 'Get started' vs 'Try free for 14 days'

PM wants to test 'Get started' vs 'Try free for 14 days'.

Put Statistical Power into practice

SaaS Product Design Tracker — From feature backlog to launch — prioritise, test, QA, ship.