How to Calculate Sample Size for a Cross-Sectional Study

“Cross-sectional study” tells us something about the design.

It does not tell us enough to choose a sample-size formula.

A cross-sectional study might aim to:

• estimate the prevalence of depression;
• estimate the proportion of facilities offering a service;
• compare contraceptive use between two groups;
• assess an exposure-outcome association;
• estimate a mean score;
• fit a multivariable regression model.

Those are not necessarily the same sample-size problem.

This article focuses first on one of the most common cases:

A cross-sectional study designed to estimate a single population proportion with specified precision.

That is the situation where the familiar Cochran-style formula is often appropriate under a simple random sampling framework.

For a large population:

Where:

• n₀ = initial required sample size
• Z = critical value for the chosen confidence level
• p = anticipated proportion or prevalence
• 1 − p = the complementary proportion
• d = desired absolute precision, often called the margin of error

For a 95% confidence level, Z ≈ 1.96.

Suppose previous credible research suggests that about 20% of the target population has the characteristic you are studying.

Use:

• p = 0.20
• 1 − p = 0.80
• Z = 1.96
• d = 0.05

This gives approximately:

n = 246

The estimate is smaller than when p = 0.50 because a binary proportion has its greatest variance around 50%, all else equal.

Do not automatically use 50% simply because it gives a familiar number. If a credible prior estimate exists and is relevant to your population and outcome, using it may be defensible.

The basic large-population calculation assumes that the target population is sufficiently large that sampling a few hundred people does not substantially deplete it.

Suppose your total eligible population is only 1,000 people.

Using the initial estimate of about 385, you can apply a finite population correction:

Where N is the population size.

With n₀ ≈ 385 and N = 1,000:

n ≈ 278

That is because observing 278 of 1,000 eligible people gives you more information about that finite population than observing 278 people from a population of several million.

Suppose the required completed sample is 278, but you expect only 90% of selected participants to provide usable data.

Divide by the anticipated response proportion:

You would need to approach approximately:

309 people

The distinction matters:

• required completed sample = 278
• number to approach = 309, given a 90% response assumption

Do not simply “add 10%” without checking what that does mathematically. Dividing by the anticipated response rate is clearer.

This is where many sample-size calculations become too optimistic.

The simple formula above corresponds to a simple random sampling framework. If you sample clusters, for example, villages, schools, facilities or enumeration areas, observations within clusters may be more similar to one another than observations selected independently across the whole population.

That reduces the effective information in the sample.

A design effect may therefore be needed when calculating the required sample size.

For example, if the simple-random-sample requirement were 385 and a justified design effect were 1.5:

You would then consider non-response and other design features as appropriate.

But do not choose a design effect because “1.5 is common.” It should be based on credible prior information or a defensible planning assumption.

And remember: if the design is clustered, the analysis must also account for the clustering. Inflating the sample size does not make a clustered design analytically simple random.

No.

This is one of the most important things to understand.

If your main objective is:

“To estimate the prevalence of hypertension with ±5% precision”

a single-proportion precision formula may be appropriate.

But if your main objective is:

“To compare hypertension prevalence between men and women”

you need a calculation suited to comparing groups.

If your objective is:

“To assess whether obesity is associated with hypertension”

you may need a power calculation based on an expected effect measure, exposure distribution and model.

If your outcome is a continuous mean rather than a proportion, a different calculation is needed.

Before opening a calculator, write down these questions:

Are you estimating one proportion, comparing groups, estimating a mean or testing an association?

Prevalence? Mean? Risk difference? Odds ratio? Another parameter?

For precision-based calculations this may be a confidence level and margin of error. For hypothesis-testing calculations, power and effect size become central.

Expected prevalence, standard deviation, anticipated effect or intracluster correlation may come from earlier studies or pilot data.

Simple random, stratified, systematic, clustered or multistage?

Non-response, missing records or attrition, depending on the design.

Only then should you calculate the number.