Is n always the number of people?

No; it can count products, accounts, trials, or other specified units.

Does more data always mean more precision?

Not necessarily, especially if the added data are redundant or the target changes.

Should planned and final sizes both be reported?

Often yes, because exclusions and response patterns help explain the evidence.

AI Glossary / Statistics and probability / Intermediate

Sample size

Definition

Sample size is the number of units or observations included in a defined sample, usually denoted n. Its meaning depends on the sampling unit, study stage, and analysis. The number of people recruited, usable measurements, independent experimental units, and records in a dataset can differ substantially, so a sample-size statement should identify exactly what is being counted.

What it means

Sample size helps describe the amount of evidence available, but it does not measure evidence quality by itself. A large collection of duplicated, biased, or highly dependent records may provide less useful information than its row count suggests.
The distinction between planned and achieved sample size also matters. A study might invite many participants, receive fewer responses, and analyze fewer still after applying predefined eligibility and completeness rules. Each count answers a different operational or statistical question.

How it works

Define the unit relevant to the conclusion. A survey about customers may count customers, while a transaction analysis may count orders. If an intervention is assigned to teams, the number of teams is important for causal uncertainty even when many employees contribute measurements.
State sample size at the appropriate stage. Report invited or selected units, respondents, excluded cases, and analyzed observations where relevant. A final total without this context can conceal substantial missingness or selection.
Distinguish total sample size from group sizes. An experiment with 200 units split equally between two groups has 100 per group, not 200 per group. Unequal allocation affects the precision of comparisons and may be deliberate.
Consider dependence. Multiple readings from the same sensor, repeated visits by one person, and observations from the same cluster may be correlated. Standard errors based on independent rows can overstate the information if that correlation is ignored.
Use effective sample size only with a stated definition. For example, weighting or clustering may reduce precision relative to a simple independent sample. An effective size is a comparison tied to an estimator and design, not a universal replacement for reporting actual counts.
Under a simple independent sampling model, many standard errors decrease approximately with the square root of n. Thus doubling the sample does not usually halve the standard error. The exact relationship can differ with finite populations, dependence, rare events, or model structure.
Choose size according to the objective. Estimating a mean to a desired precision, detecting a meaningful effect with adequate power, and training a complex predictor are different tasks. There is no single minimum sample size that makes every analysis valid.
Review coverage and quality alongside quantity. Larger samples do not automatically remove confounding, measurement error, nonresponse bias, or distribution shift. In AI evaluation, sufficient overall size may still leave too few cases for a small but important subgroup or rare failure mode.

Three practical examples

A/B experiments: researchers report both total assigned users and counts in each variant.
Sensor studies: analysts distinguish the number of devices from the number of repeated readings.
AI assessment: a team reports total test cases and the separate counts supporting each subgroup result.

Think of it this way

Think of evidence as observations collected from sources. Asking one source the same closely related question many times does not necessarily provide as much independent information as asking many independent sources. The sample-size label needs to reveal both the count and what the counted units represent.

Worked example

A fictional monitoring study collects five readings from each of 200 devices. It contains 200×5 = 1,000 measurement records, but only 200 distinct devices. Whether the readings can be treated as independent depends on the process; the row count alone cannot establish that.
Separately, consider a simple illustrative mean-estimation model with independent identically distributed observations and a known population standard deviation of 12 units. With n = 100, the standard error of the sample mean is 12/sqrt(100) = 1.2 units.
With n = 400 under the same model, it is 12/sqrt(400) = 0.6 units. Quadrupling the sample halves this standard error. The comparison illustrates the square-root relationship, not a guarantee for the repeated-device data or for every sampling design.

Illustrative case study

Fictional case: Meadow Analytics evaluates an AI assistant on 50,000 logged interactions. The dashboard presents the count as overwhelming evidence, but a review finds that most records come from a small set of unusually active accounts.
The team reports distinct-account counts, examines within-account dependence, and adds coverage of less active customers. It retains the large interaction total as a descriptive count without treating it as 50,000 independent customer experiences.

Common misconceptions and mistakes

A large n does not guarantee representativeness or eliminate bias.
Thirty observations is not a universal threshold that makes every distribution normal or every analysis trustworthy.
The number of records is not always the number of independent units.
A study with a large overall sample can remain weak for rare outcomes or small subgroups.
A tiny p-value in a huge sample can correspond to an operationally trivial effect, so magnitude and uncertainty still matter.

Business use

Use sample size as a clearly defined part of study documentation and planning. Before accepting a performance claim, ask what n counts, how cases were selected, whether observations are dependent, and how many cases support the particular comparison of interest. These questions often matter more than the largest number on a presentation slide.

Check your understanding

Answer: An experiment has 300 users split 2:1.

How many are in each group?

Answer: Two hundred and 100.

Under the simple independent mean model, how much must n increase to halve the standard error?

Answer: By a factor of four.

Why are 1,000 device readings not automatically 1,000 independent units?

Answer: Several readings come from each device.

Related terms

Learn next

Audit a study's counts from selection to analysis. Write separate totals for unique units, repeated records, treatment groups, and important subgroups, then identify the assumptions used to calculate uncertainty.

Sources

NIST/SEMATECH e-Handbook, Sample sizes required for a mean: https://www.itl.nist.gov/div898/handbook/prc/section2/prc222.htm OECD Glossary of Statistical Terms, Sample size: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf The device counts and precision comparison are original illustrations.

Previous: Sample rotation Next: Sample size determination