Should the number be rounded upward?

Yes, when meeting a calculated minimum count.

Does a larger population always require a proportionally larger sample?

No; the relationship depends on the design and sampling fraction.

Can the plan change during the study?

Sometimes, but adaptive or revised plans need appropriate rules and transparent documentation.

AI Glossary / Statistics and probability / Intermediate

Sample size determination

Definition

Sample size determination is the process of choosing how many units or observations a study needs to meet a defined objective under stated assumptions. The objective may concern estimation precision, statistical power, model performance, or another criterion. A defensible calculation connects the number to the design, expected variability, meaningful effect or error tolerance, and practical losses rather than relying on a universal rule of thumb.

What it means

The question is not simply how many records can be collected. It is what evidence the study must provide and how much data are needed to support that evidence with acceptable uncertainty.
A precision calculation asks how narrow an interval should be. A power calculation asks how likely a test is to detect a specified effect when it exists. These purposes can yield different sample sizes and should not be presented as interchangeable guarantees.

How it works

Define the primary estimand, outcome, population, and analysis before calculating size. Specify whether the study estimates one average, compares groups, evaluates a proportion, or measures a rare event. Multiple primary objectives may require additional planning.
For precision, state the desired interval coverage and margin of error in meaningful units. In a simple independent normal model with known standard deviation sigma, a two-sided approximate 95% mean interval has half-width 1.96 sigma/sqrt(n). Solving for n gives a starting requirement.
For hypothesis testing, specify the significance level, target power, effect size of practical interest, allocation ratio, and variability. Power is evaluated at a particular alternative, not for every possible effect. Selecting an unrealistically large expected effect can make a small study appear adequate.
Use defensible planning inputs from relevant prior data, pilot work, or plausible scenarios. Estimated variances and event rates are uncertain. Examine several plausible values rather than treating a fragile pilot estimate as a known constant.
Account for the sampling and assignment design. Clustering, repeated measurements, unequal weights, stratification, and finite-population sampling can change the information contributed by each unit. A simple independent-observation formula may be inappropriate without modification.
Allow for anticipated nonresponse, dropout, or unusable measurements. An inflation based on an expected retention rate is a planning approximation, not a guarantee that the required final count will be achieved. Differential loss can also introduce bias that a larger invitation count does not remove.
Round required counts upward and distinguish total, per-group, per-cluster, and evaluable counts. Check whether enough independent clusters or events are present, not merely enough rows. Operational constraints should be made explicit rather than hidden inside a nominal total.
For complex models, simulation may better represent the design and planned analysis. Validate the simulation assumptions and code. Prespecify stopping or adaptive rules so a study does not repeatedly inspect ordinary p-values and stop only when a desired result appears.

Three practical examples

Customer research: a survey is sized to estimate an average score within a defined margin.
Product testing: an A/B experiment is planned to detect an operationally meaningful change with specified power.
AI validation: a team plans overall and subgroup evaluation counts to estimate performance with useful uncertainty.

Think of it this way

Think of choosing the resolution needed in a measurement before buying the measuring effort. A broad outline requires less information than a fine distinction. The calculation translates the desired resolution and the expected noisiness of the situation into a collection target, subject to its assumptions.

Worked example

Suppose a fictional study wants an approximate 95% confidence interval for a population mean with half-width at most 3 units. Assume independent observations and use a planning standard deviation of 15 units.
The normal-model calculation is n ≥ (1.96×15/3)² = 96.04. Rounding upward gives 97 evaluable observations. At n = 97, the calculated half-width is approximately 2.985 units.
If 80% of recruited units are expected to yield usable data, a simple recruitment target is ceiling(97/0.80) = 122. This targets an expected usable count; it does not guarantee at least 97 completions.
As a sensitivity check, using standard deviation 18 instead produces n ≥ (1.96×18/3)² = 138.2976, so 139 evaluable observations are needed. If sigma is estimated rather than known or the normal approximation is unsuitable, a more appropriate interval and planning method are required.

Illustrative case study

Fictional case: Alder Support plans to evaluate average handling time after a workflow redesign. The initial proposal uses a convenient recruitment target without specifying precision.
The analyst defines the target outcome, checks historical variability, and calculates requirements under several plausible variance and retention scenarios. Because work is assigned by team, the final plan also accounts for clustering. Management receives a clear account of what the feasible study can estimate, rather than a promise attached to an arbitrary sample count.

Common misconceptions and mistakes

There is no universally adequate sample size such as 30, 100, or 1,000 for every purpose.
A power calculation is not a prediction that the study will produce a significant result.
A margin of error does not include every source of bias or measurement error.
Inflating for attrition does not correct selective missingness.
A calculation based on independent people cannot be transferred unchanged to a small number of correlated teams.
Retrospective power based only on the observed effect adds little beyond the observed estimate and uncertainty.

Business use

Use sample size determination to align data collection costs with the decisions the evidence must support. Document the objective, assumptions, formulas or simulation, sensitivity scenarios, and final feasible target. If the available budget cannot achieve the intended precision, make that limitation visible and consider refining the question or design.

Check your understanding

Why does the worked example use 97 rather than 96?

Answer: The calculated requirement is 96.04, so 96 falls short.

Does recruiting 122 guarantee 97 usable observations at an expected 80% retention rate?

Answer: No; realized retention varies.

What happens when the planning standard deviation increases?

Answer: More observations are needed for the same margin.

Related terms

Sample size
Effect size
Design effect
Confidence interval
statistical power
margin of error
attrition
simulation

Learn next

Create a sensitivity table varying the expected variability, target effect or precision, and retention rate. Check whether the planned analysis matches the assumptions used for each calculation.

Sources

NIST/SEMATECH e-Handbook, Sample sizes required for inference about a mean: https://www.itl.nist.gov/div898/handbook/prc/section2/prc222.htm The precision, recruitment, and sensitivity calculations and business case are original illustrations.

Previous: Sample size Next: Sample size disparity