01Locate the level at which the intervention was assigned

NC3Rs defines the experimental unit through independent exposure to an intervention and warns that misidentifying it can inflate the apparent sample size. Its examples show that the relevant unit need not always be an individual animal.NC3Rs EDA — Experimental unit (opens in a new tab)

The same reading principle matters in cell and tissue work. A donor, culture preparation, well, plate or experimental run can occupy a different level in the design. The level relevant to one factor may differ from the level relevant to another.

A label such as triplicate does not answer these questions. Three instrument readings, three wells and three independently sourced cultures can provide different information.

02Count readings and units separately

In an original hypothetical experiment, three independently treated culture preparations each produce four repeated instrument readings. The dataset contains twelve readings, but the condition comparison has three independently treated preparations in that group.

Reporting n = 12 without explaining the nesting can make the biological evidence appear larger than it is. The repeated readings help characterise measurement variation within each preparation; they do not add nine new independently treated preparations.

Averaging the four readings within each preparation may be appropriate in a simple design, while a suitable hierarchical analysis can retain their structure. The choice depends on the design and analysis question rather than a universal instruction to average everything.

03Do not infer independence from separate containers

Separate wells may share a donor, original culture, preparation batch or plate environment. Those shared features can matter to the inference even when the well-level intervention was applied independently.

NC3Rs notes that one biological source can sometimes provide several experimental units, depending on the intervention, while generalisation beyond that source remains limited. Shared-source structure may need to be reflected in design and analysis.NC3Rs EDA — Experimental unit (opens in a new tab)

This prevents the opposite oversimplification: declaring every measurement from one source to be merely technical regardless of how treatments were assigned. Experimental independence and biological generalisability are related but distinct questions.

For example, independent well-level treatments from one donor can inform a comparison within that donor’s material. They do not establish how consistent the result is across donors, because donor-to-donor variation was not observed.

04Read n, error bars and exclusions together

Look for a description of independent experiments, source material, subsampling and repeated measurements. Then check what each plotted point and error bar represents: individual readings, unit means or a model-derived summary.

An error bar based only on repeat instrument readings can be narrow while the biological units vary substantially. Its interpretation depends on the level of variation included, not just its visible width.

Record exclusions at the correct level. Removing one failed reading from a preparation is different from excluding that preparation entirely; the remaining unit count and analysis should reflect the actual decision.

If a paper leaves the replicate structure unclear, state that limitation rather than guessing from the number of dots. The most useful evidence summary names the independent units and the additional readings that support their measurement.