A small sample, a narrow conclusion: reading animal cognition studies

How many animals were behind the claim?

A small sample means a narrow conclusion: a cognition finding describes the animals actually tested, not the whole species. A headline about clever crows or empathetic rats rarely tells you how many animals produced the result. That number sets how far the finding can reach. A study with six subjects and a study with sixty are not making claims of the same weight, even if both report a “significant” effect.

Sample size sits second in the ARRIVE guidelines’ Essential 10, right after study design. The guidelines describe these ten items as “the basic minimum that must be included in any manuscript describing animal research,” adding that “without this information readers and reviewers cannot assess the reliability of the findings.” If a paper or a press summary skips the count, there is no way to judge whether the effect is stable or a product of a handful of unusually responsive individuals.

Counting animals sounds mechanical, but it is where a lot of cognition reporting quietly breaks down. A study might describe “subjects” or “participants” without separating individuals from sessions, trials, or measurements taken repeatedly from the same few animals. Those are different kinds of numbers, and conflating them inflates the apparent evidence base.

Before accepting a claim about problem-solving, tool use, or memory in a species, ask a short set of questions:

• How many distinct animals were tested, separate from how many data points were collected?
• Does the reported number match the number described in the methods, or only the number in the abstract?
• Were any animals tested more than once under different conditions, and is that made clear?

None of this requires statistical training. It requires reading past the summary sentence to the section where the animals themselves are described, and treating that count as the boundary of what the study can say.

Separate an animal from repeated trials

The animal, not the trial, is usually the unit a cognition claim is allowed to generalize about. This distinction gets blurred constantly in casual science writing, where “100 trials showed strong learning” can quietly stand in for a study that used four animals run twenty-five times each.

The NC3Rs experimental design resource defines the experimental unit as “the entity you want to make inferences about (in the population) based on the sample (in your experiment),” and adds a direct warning: “if you take multiple measurements from the same animal it does not mean that each animal provides multiple experimental units.” The unit is defined by what receives the intervention. How many times it gets measured afterward is a separate count.

That distinction matters because the same resource notes that failing to identify the experimental unit correctly risks overestimating sample size, which “could invalidate the results of your statistical analysis and conclusions.” A memory test run one hundred times on four animals is a study of four experimental units with repeated measurements. Four is the sample size, and one hundred is the count of measurements.

The unit itself is not always the whole animal, and the NC3Rs page works through the cases. A pregnant female treated with measurements taken on her pups. A cage of animals where the treatment reaches everyone in it. A patch of skin, where topical application divides one animal’s back into several independently treated units. An animal for a period of time, in a crossover design where each animal is its own control across washout periods. Each of those changes what the reported sample size actually counts.

For a reader trying to weigh a cognition finding, the practical move is to locate two numbers in the methods section: the count of animals, and the count of trials or measurements per animal. When only the second number is reported prominently, treat the claim as resting on fewer independent data points than the headline volume suggests. The gap between “measurements taken” and “animals tested” is where overconfident interpretation usually starts.

Which animals did not reach the analysis?

An animal enrolled in a study and an animal counted in the final analysis are not always the same animal, and that gap deserves attention before a finding gets generalized. Cognition studies with live subjects can lose data points for reasons that have nothing to do with the ability being tested: an animal disengages from a task, a session is excluded for equipment failure, or a subject fails to meet a pre-set training criterion and never reaches the test phase.

The ARRIVE Essential 10 lists inclusion and exclusion criteria as its own reporting item, separate from sample size and separate from study design. That separation reflects a simple concern: a reader assessing reliability needs to know how many animals started, how many were dropped, and why.

When exclusions are not reported clearly, a study can look like it tested a broader group than it did. An abstract describing “twelve subjects” might reflect twelve animals that completed every phase, after an unreported number of others were removed earlier for reasons the reader never sees. That is a different claim than one built on twelve animals with a transparent account of who did not make it into the final numbers.

For anyone reading a cognition paper, or a summary of one, it is worth checking the methods for a specific accounting:

• The number of animals originally enrolled or bred into the study.
• The number retained at each stage, from training through final testing.
• The stated reasons behind each exclusion, beyond a single final count with no history attached.

Where that accounting is missing, the analyzed sample cannot be assumed to represent the full group of animals the researchers started with. That absence is itself a piece of information worth noting. Recording the gap matters more than guessing at what filled it.

Carry the sample boundary into the conclusion

A cognition finding applies to the animals actually measured, and stretching it past that boundary is where interpretation outruns evidence. Once you know how many individuals were tested, how the experimental unit was defined, and which animals made it into the final analysis, the next step is checking whether the study’s own conclusion respects those limits.

The NC3Rs experimental design resource makes the stakes concrete in one direction. Where a single animal can legitimately supply many experimental units, it warns that using only one animal keeps the results specific to that animal. Establishing a population-level result takes more than one. It also gives the remedy in the same breath: use the individual animal as a blocking factor and use more than one animal to improve generalisability. A striking result from one subject, however many times it was tested, describes that one subject.

ARRIVE addresses the other end of this problem, but notably not in its minimum set. Interpretation and scientific implications is item 17 and generalisability and translation is item 18, and both live in the Recommended set, separate from the Essential 10. The guidelines describe that second set as complementing the first and adding important context, with reporting both representing best practice. Check those Recommended items explicitly when judging how far a conclusion travels. Their placement in the guidelines does not tell you what a particular journal enforces.

Practical reading habits follow from this. Treat a species-wide claim (“dogs understand pointing gestures”) differently from a sample-specific finding (“the tested group of dogs responded to pointing gestures in this task”).

Watch for conclusions that shift from describing what was measured to asserting what the behavior means for the species, a household pet, or animal welfare broadly, since that shift is an interpretive step the data alone cannot make.

Organizations summarizing cognition research for a general audience can adopt, as an editorial choice, the habit of stating the animal count and exclusion history alongside any headline finding.

The sample boundary drawn in the methods section is the boundary the conclusion should respect, and checking that alignment is the reader’s job as much as the author’s.