Inference

Sample size for estimating a proportion

Consider planning a basic proportion estimate for the precision you desire. When selecting your option for calculations, distinguish study power from survey precision.

Define the planning task

Population size alone isn't enough information to provide a single minimum sample size. The process of finding an estimated population percent for a certain degree of precision is different from detecting the difference between treatments, estimating regression coefficients, or discovering rare side effects.

The plan below provides a calculation for the sample size needed for an estimate of one population proportion using simple random sampling. Precision is considered, but not clinical trial power. It does not automatically address issues of clustering, multiple measurements, weights or response rates.

Choose confidence, proportion and precision

Selecting a confidence level and its z critical value, proportion (p) and precision (E) will depend on your research question. For large populations, you could use the following equation: n = z^2*p*(1-p)/E^2. In this case p would be the anticipated proportion of interest. E would be the margin of error as a decimal. A good example of an estimate with a margin of 5 percentage points is E = 0.05.

Setting p to .50 would maximize the value of p(1 − p) and thus would give you the greatest target value possible within this equation when no reasonable estimate is available. Making an educated estimate may reduce your target value. However, a poor estimate can result in less precision than you originally intended. Always round up to the nearest whole number.

Account for the actual design

If you are sampling without replacement from a known finite population, a finite-population adjustment may need to be made. Whether a finite-population adjustment should be used is dependent upon both the design of the sampling method and the population size. Knowing the population size is necessary, but not sufficient to warrant its use.

Target complete usable responses. For instance, if you have determined you need to collect at least 385 complete usable responses and anticipate a 70% usable response rate, then you will need to invite at least 550 participants. Even so, there can still be respondent bias. Additionally, there is much additional planning needed for the power of a comparison (i.e., effect size, variability or risk as applicable to your situation, alpha level and study design).

Worked example: a survey percentage

Calculate an estimate of a single population proportion with 95% confidence and with a margin of five percentage points or 0.05. Set p = .50 and assume a large-population approximation.

n = 1.96² × 0.5 × 0.5 / 0.05² = 384.16; round up to 385 completed observations.

The above is an approximate calculation for a simple-random-sampling scenario providing a precision target. It doesn’t guarantee an unbiased sample, nor does it calculate participant numbers for comparing treatments.

Determine whether you are after precision or statistical power first. Next determine which of the formulas fits the goals of your study and design of your study.

Sources for further reading