Thematic saturation calculator

Saturation is the most common justification for a qualitative sample size and the least often demonstrated. Morse's complaint from 1995, that saturation is "evident mainly by declaration", still describes most methods sections. The tool above turns the declaration into a reported measurement, and the planning tab covers the earlier question every ethics committee asks: how many participants, and why.

The Guest, Namey and Chen method

Guest, Namey and Chen (2020, PLOS ONE) proposed three parameters that make a saturation claim specific. The base size is how many initial interviews define the themes already known; they tested 4, 5 and 6. The run length is how many consecutive further interviews are examined for new themes, 2 or 3. The new-information threshold is how much novelty a run may add and still count as saturated: ≤5% or 0% of the base's themes.

Their worked example makes the arithmetic concrete. The first four interviews produce 37 themes. Interviews 5 and 6 add seven, a ratio of 19%; interviews 6 and 7 add four, 11%; interviews 7 and 8 add one, 2.7%, which is under the 5% threshold. Saturation is therefore reported at 6⁺²—reached at interview 6 and confirmed by two more. At the stricter 0% threshold the same data saturate at 10⁺². That example is loaded in the tool by default, reconstructed from every figure the paper states.

Two properties make the method more defensible than older approaches. The denominator is the base, not the whole dataset, so saturation is not guaranteed by construction the way it is when new themes are expressed as a share of all themes eventually found. And it works prospectively: you can compute it after each interview and stop recruiting when the threshold is met, rather than only after analysis is finished.

Report the setting, not just the word

Saturation at 6⁺² with a 5% threshold is a weaker claim than 10⁺³ at 0%, and readers can only weigh it if they are told which one it is. The grid in the tool shows all four of the paper's reporting options for your data, and the strictest one you meet is the most persuasive to report. State the base size too: a large base makes the ratio smaller and saturation easier to reach.

Order matters, so test it

The method counts themes in the order interviews were analysed. A dataset where the most talkative participants happened to come first will saturate early; the same interviews in a different order may not. Guest and colleagues addressed this by bootstrapping random orderings of three real datasets. With a theme × interview matrix—the code-by-document matrix NVivo, ATLAS.ti and MAXQDA all export—the tool repeats the assessment on 1,000 seeded random orders and reports how often saturation arrived as early as it did in yours. A figure above 80% is reassuring; a figure near 50% means your saturation point is partly an accident of scheduling.

Before data collection: justifying the number

Three published aids are worth citing in a protocol, and the planning tab combines them.

• Empirical benchmarks. Hennink and Kaiser's (2022) systematic review of 23 studies found saturation within 9–17 individual interviews or 4–8 focus group discussions—mostly for homogeneous samples with narrow objectives, and mostly for code saturation, which is reached earlier than the fuller "meaning saturation".
• Information power. Malterud, Siersma and Guassora (2016) argue that the more relevant information a sample holds, the fewer participants it needs, and name five dimensions that decide it: the breadth of the aim, the specificity of the sample, the use of established theory, the quality of dialogue, and case versus cross-case analysis. The model deliberately gives no formula, so the tool does not invent one; it turns your ratings into the argument.
• A binomial bound. Fugard and Potts (2015) compute how many participants are needed to observe a theme of a given prevalence a given number of times. Their example—29 participants for an 80% chance of seeing a theme held by 10% of the population at least twice—is reproduced exactly. It assumes random sampling, which purposive qualitative samples are not, so it is a sanity check on the order of magnitude rather than a requirement.

When saturation is the wrong criterion

Saturation assumes that themes are discovered in the data and that, after enough interviews, there are no more to find. That fits codebook and coding-reliability approaches, where a stable set of codes is applied across transcripts, and it is the assumption behind every number on this page. It fits reflexive thematic analysis poorly: Braun and Clarke argue that themes are generated by the analyst's engagement with the data, so there is no fixed set to run out of. If that is your approach, use the information power argument and leave the saturation metric out.

Saturation also says nothing about whether themes were coded consistently. If two people applied the codebook, the intercoder reliability calculator pools their agreement from the same NVivo, ATLAS.ti or MAXQDA export, and the codebook generator helps build the definitions that make codes applicable the same way twice. For quantitative reliability studies, where the question is how many subjects are needed to estimate kappa or an ICC precisely, the reliability sample-size planner is the right tool instead.

What this tool does not do

It counts themes, not their depth: two interviews that each mention a theme once contribute the same as one that explores it at length, which is why code saturation arrives earlier than meaning saturation. It trusts the order you give it, so paste interviews in the order they were analysed, not alphabetically. And it cannot tell you whether a theme is important; a single new theme that changes the answer to your research question matters more than any ratio.

Frequently asked questions