Sample size determination is the process of deciding how many observations are needed before collecting data. It matters because too small a sample can give results that are too uncertain, while too large a sample can waste time, money, or participants. In estimation problems, the goal is often to achieve a chosen margin of error at a chosen confidence level.
In hypothesis tests or experiments, the goal is often to have enough power to detect an important effect.
Understanding Statistics: Sample Size Determination
A sample does not become reliable simply because it seems large. Its usefulness depends on how much people, objects, or measurements differ from one another. A class might have very similar heights, while household incomes in a town can vary greatly.
Greater natural variation requires more observations to pin down an average. The key pattern is based on square roots.
To make typical sampling uncertainty half as large, a study usually needs about four times as many independent observations. This is why collecting a few extra results has little effect in a very small study, while large improvements in precision can become expensive.
Planning starts with assumptions, not certainties. Researchers often use earlier studies, a pilot study, or subject knowledge to estimate how spread out the data will be. A poor estimate can lead to a poor target size.
For a survey percentage, the hardest case is often a result near one half, since responses are then most mixed. Results close to zero or one tend to need fewer responses for the same precision.
If the whole population is small, such as every student in one year group, sampling a large share of that population gives more information than the usual large-population calculation suggests. This adjustment is called the finite population correction.
The way observations are collected can matter more than the raw count. One thousand answers from a voluntary online poll may be less useful than two hundred randomly selected answers, because volunteers can differ from people who do not respond. Responses from pupils in the same classroom, patients at the same clinic, or trees in the same field are often similar.
They do not provide as much separate information as truly independent observations. This is clustering. Clustered designs commonly need a larger total sample.
Missing data matters too. If some participants are likely to leave a study, researchers recruit extra people at the start so that enough usable results remain.
For an experiment, the required size depends on the smallest difference worth detecting in real life. A tiny difference may be statistically detectable with a huge study, yet still be unimportant. A large useful difference can be found with fewer participants.
Researchers must choose this practical effect before seeing the results. They must account for measurement noise, the chance of a false positive result, and the desired chance of finding a real effect.
Using matched pairs or repeated measurements can reduce noise when each person is compared with themselves. Good sample size work states every assumption clearly, checks whether the planned design matches those assumptions, and treats the final number as a justified estimate rather than a guarantee.
Key Facts
- For estimating a population mean with known or planned standard deviation: n = (z*σ / E)^2
- For estimating a population proportion: n = z*^2 p(1 - p) / E^2
- If p is unknown for a proportion, use p = 0.50 because it gives the largest required sample size.
- Margin of error for a mean is E = z*σ / sqrt(n), so quadrupling n cuts E in half.
- For comparing two equal-size group means, a planning formula is n per group = 2(zα/2 + zβ)^2 σ^2 / Δ^2
- Power is the probability of detecting a real effect, commonly set to 0.80 or 0.90 before the study begins.
Vocabulary
- Sample size
- The number of observations, measurements, or participants included in a study.
- Margin of error
- The maximum expected distance between a sample estimate and the true population value for a stated confidence level.
- Confidence level
- The long-run percentage of confidence intervals made by the same method that would contain the true population value.
- Statistical power
- The probability that a test correctly rejects a false null hypothesis when a real effect exists.
- Effect size
- The size of the difference or relationship a study is designed to detect, often written as Δ for a difference in means.
Common Mistakes to Avoid
- Using n instead of sqrt(n) in the margin of error formula is wrong because uncertainty decreases with the square root of sample size, not directly with sample size.
- Forgetting to round the final sample size up is wrong because a fraction of an observation cannot be collected and rounding down may miss the target precision or power.
- Using p = 0.50 only after data are collected is wrong because p = 0.50 is a conservative planning value used before the true population proportion is known.
- Confusing confidence level with power is wrong because confidence describes interval coverage for estimation, while power describes the chance of detecting a real effect in a test.
Practice Questions
- 1 A researcher wants to estimate a population mean with σ = 12, margin of error E = 3, and 95% confidence. Use z* = 1.96 to find the required sample size, rounding up.
- 2 A pollster wants a 95% confidence interval for a population proportion with margin of error E = 0.04 and no prior estimate of p. Use p = 0.50 and z* = 1.96 to find the required sample size, rounding up.
- 3 A study team can either increase confidence from 95% to 99% or decrease the desired margin of error while keeping the same population variability. Explain how each choice affects the required sample size and why.