How many participants do you need for a reliable picture

2026-07-03

The question comes up in every project: how many people have to take part before you can take the result seriously?

The answer is more reassuring than expected, and it turns out to be about something other than numbers.

From about ten, comfortable from fifteen

Group Concept Mapping already works from around ten participants. From fifteen to forty it comes into its own. Above that the picture barely becomes more stable; you are mostly adding confirmation.

That is because the map does not rest on individual opinions but on similarities in how people group things. Patterns like that stabilise faster than averages do.

How you know it was enough

You do not have to judge this by feel. Split-half reliability repeatedly splits participants into two random halves and compares the maps each half produces. If those correspond strongly, the outcome does not depend on who happened to take part.

Above 0.80 counts as sound; the average in the published literature sits around 0.87. That figure appears in every report, so you can always check it.

Not: how many people took part. Rather: would a different half have produced the same map.

Spread weighs more heavily than number

Thirty people from one department produce a narrower picture than fifteen people from five parts of the organisation. If you later want to compare groups, teachers against school leaders for example, you also need enough people per group; below eight such a comparison becomes shaky.

Not everyone has to do everything

A practical point that gives a lot of room: the phases may have different groups. Contributing ideas can be broad, sorting can be done by a smaller group, rating broad again. Sorting is the most demanding part, so it is fine to ask a selection for that as long as the selection reflects the breadth.

And if there are too few

Below ten completed sortings the map becomes restless: small shifts in who took part then visibly change the picture. You can still run the analysis, but present the outcome as an exploration rather than as a shared view.

More on how reliability is calculated is on the method page.