More sample can mean more material from one vial, more vials from a lot, or more readings from the same prepared solution. These changes provide different kinds of information. Confidence in a batch assessment depends on the sampling design and the question, not simply on a larger number in a report.
Identify what the sample count represents
| Reported number | What it counts |
|---|---|
| 20 mg submitted | Material quantity, possibly from one unit |
| 20 vials selected | Original units, if separately identified |
| 20 injections | Instrument readings, possibly from one solution |
| 20 aliquots | Portions whose independence depends on their origin |
A laboratory may need enough material for an analytical method. That requirement is distinct from choosing enough independent units to investigate a batch. Increasing the amount from one vial can help the first task without addressing the second.
Keep the count connected to selection. Twenty deliberately chosen attractive vials and twenty randomly selected vials are not equivalent evidence about a lot, despite matching counts.
Work through a conditional detection example
NIST describes a binomial approximation for random samples when the lot is large relative to the sample. It links sample size and a specified defective fraction to the probability of observing defects.NIST — Choosing a Sampling Plan with a given OC Curve (opens in a new tab)
For an original illustration, suppose 5% of a very large lot have a particular defect. Assume independent random selections and a test that detects that defect perfectly. The probability of seeing none in n selected units is 0.95 raised to n; the probability of detecting at least one is one minus that value.
| Units selected | Chance of detecting at least one |
|---|---|
| 5 | 22.6% |
| 20 | 64.2% |
| 60 | 95.4% |
The table describes what repeated sampling would detect if the assumed defective fraction were true. It is not an observed Novum defect rate, a recommended sampling plan or a statement that a particular lot has a 95.4% chance of being satisfactory.
Know when the simple calculation stops fitting
The NIST approximation assumes the lot is sufficiently large that sampling does not materially change what remains. If a substantial fraction of a finite lot is sampled without replacement, an appropriate finite-population calculation is needed.NIST — Choosing a Sampling Plan with a given OC Curve (opens in a new tab)
Clustering also matters. Units taken from one local group may share a problem or all miss a problem confined elsewhere. Treating strongly related selections as independent can overstate the information gained.
The assumed perfect test is another simplification. A method that misses some affected units changes the detection probability. A test for one defect says nothing automatic about defects that it does not measure.
Finally, detecting any failure and estimating a mean content are different statistical objectives. A sample size selected for one purpose cannot be justified for the other by copying the same probability table.
Ask what improvement another sample would provide
A sound assessment starts with the property of concern, the population to be described and the consequence of an incorrect conclusion. Those choices determine what information is missing and how additional units could help.
If uncertainty comes mainly from an unclear measurement basis, more vials analysed on that same unclear basis may not solve the problem. If the main gap is coverage of different parts of the lot, additional independent selections may be more informative than more injections.
Report the sample count alongside its rationale and limitations. A bare claim of statistical confidence is difficult to evaluate when the underlying question, model and selection process are absent.
Sources and further detail
- NIST — Choosing a Sampling Plan with a given OC Curve (opens in a new tab)
Official binomial sampling derivation read. The 5%, n=5/20/60 detection illustration is calculated independently under explicit assumptions and is not an acceptance recommendation.
Sources checked 19 September 2026. Worked examples are illustrative unless a supplied report is explicitly identified. This article has not undergone independent scientific peer review.