How many participants do I need? A quick power-analysis primer

August 5, 2026 · 6 min read

“How many participants do I need?” is the most common question in study design, and the honest answer is: it depends on how big your effect is and how sure you want to be that you'll detect it. Here's the plain-English version.

The three levers

• Effect size— how big the difference you're looking for is. Smaller effects need bigger samples. This is the lever people underestimate most.
• Power— your chance of detecting a real effect. Convention is 80%, but 90% is safer for effects you can't easily re-run.
• Alpha — your tolerance for a false positive, usually 5%.

Rough rules of thumb

For a simple two-group comparison at 80% power and a medium effect (d ≈ 0.5), you need roughly 64 people per condition. Halve the effect (d ≈ 0.25) and that jumps to about 250 per condition. The takeaway: small effects are expensive, and guessing your effect size is the biggest source of underpowered studies.

Avoid the underpowered-pilot trap

A common mistake is running a tiny pilot, seeing a “promising” result, and scaling up — when the pilot was too small to mean anything. Pilots are great for checking that a study works; they're unreliable for estimating how big an effect is.

Where AI pretesting fits

You can't derive your final sample size from AI respondents — but you can use a cheap AI pretest to sanity-check direction and catch a manipulation that does nothing before you commit to a power analysis at all. Use tools like G*Powerfor the formal calculation, and use a pretest to make sure the thing you're powering for is actually there.