How to Determine the Right Sample Size for Your Research

Ask any researcher about the toughest early-stage decision in a study, and “how many people should I survey?” will almost certainly make the list. Pick too few, and your findings lose credibility. Pick too many, and you burn through time, money, and goodwill for marginal gains. Sample size sits at the heart of every credible piece of empirical work, from a district-level public policy evaluation to a national household survey. Getting it right is less about memorising formulas and more about understanding what each number in that formula represents and why it matters.

What sample size really means 🔗

Sample size, usually written as n, is simply the number of units you observe or survey from a larger population. It directly shapes two statistical qualities of your study: the precision of your estimates and the power of the study to detect real effects. A well-chosen sample lets you say something meaningful about a population of millions by studying only a few hundred people. A poorly chosen one produces numbers that look scientific but cannot be trusted.

The key idea is representation. You are using a smaller group to stand in for a much larger one, so the sample must reflect the population’s characteristics closely enough that your conclusions hold. If the sample is too small or skewed, you may end up with findings that neither replicate nor generalise.

The factors that drive sample size 🔗

There is no universal “correct” number. The appropriate size emerges from a few interacting factors, each of which a researcher must think through carefully before collecting a single response.

Confidence level 🔗

The confidence level tells you how sure you want to be that your sample estimate falls within a specified range of the true population value. Most social science and public administration research uses a 95% confidence level, which corresponds to a z-score of 1.96. For high-stakes medical or pharmaceutical work, researchers often move to 99% confidence, which uses a z-score of 2.576. Raising confidence from 95% to 99% increases the required sample size noticeably because you are demanding more certainty from your data.

Margin of error (precision) 🔗

The margin of error, sometimes called the confidence interval, is the maximum acceptable difference between your sample estimate and the true population value. It is the cushion of uncertainty you are willing to live with. A tighter margin, such as ±2%, demands a much larger sample than a looser ±5%, holding confidence constant. This relationship is not linear – halving the margin of error roughly quadruples the required sample.

Population heterogeneity 🔗

How different are the units in your population from one another? If you are studying a highly diverse group – say, opinions across urban and rural households in a large state – the variability is high and you will need a bigger sample to capture that diversity. A more homogenous population, in contrast, can be accurately represented with fewer observations. Heterogeneity is typically captured through the population standard deviation (σ) when estimating a mean, or through the expected proportion (p) when estimating a percentage.

Frequency of the trait being examined 🔗

When your research focuses on a proportion – say, the share of citizens aware of a government welfare scheme – the expected frequency of that trait matters. If you have no prior idea about the proportion, the conservative choice is p = 0.5, because that value maximises variability and yields the largest required sample. The product p(1−p) reaches its mathematical maximum at 0.5, so using this default protects against underestimating the needed sample. If pilot data suggests the trait is rare (say, 10%) or very common (say, 90%), the required sample size drops.

Population size 🔗

Interestingly, for large populations, the total size matters far less than people assume. A sample of roughly 385 is enough to draw inferences for nearly any large population at a 95% confidence level with a 5% margin of error, which explains why samples of 400 or 500 appear again and again in published research. For smaller, finite populations, though, a correction factor does reduce the needed sample.

The role of the standard error 🔗

Standard error is the bridge between sample size and precision. It is the standard deviation of the sampling distribution – essentially, how much sample estimates would vary if you repeated your study many times. The standard error shrinks as sample size grows, because larger samples smooth out the effects of random variation.

For estimating a mean, standard error (SE) equals σ divided by the square root of n. For a proportion, SE equals the square root of p(1−p)/n. Once you know the standard error you are willing to tolerate for a given precision, rearranging the formula gives you the required sample size. This is why the estimation process always begins with thinking about precision first, then working backward to find n.

The two core formulas 🔗

Most public administration and social research ultimately boils down to estimating either a mean (like average monthly income of beneficiaries) or a proportion (like percentage of citizens satisfied with a service). Each has its own standard formula.

Sample size for estimating a proportion 🔗

The most widely used formula in survey research comes from William G. Cochran’s classic 1977 text Sampling Techniques. For a large population, the minimum sample size is:

n₀ = Z² × p(1−p) / e²

Here, Z is the z-score matching your confidence level, p is the expected proportion, and e is the margin of error. Plugging in Z = 1.96, p = 0.5, and e = 0.05 gives the familiar result of 384 – the number you will see cited in countless survey reports.

Sample size for estimating a mean 🔗

When the outcome is a continuous variable rather than a yes/no response, the formula shifts slightly. For studies measuring means, the formula becomes n = Z²σ²/e², where σ is the estimated population standard deviation. Here, σ usually comes from a pilot study, prior literature, or a reasonable estimate of the range of likely values.

Finite population correction 🔗

If your population is small – say, all 528 employees of a state department – the large-population formula overshoots. A finite population correction brings the number down to something more practical:

n = n₀ / [1 + (n₀ − 1)/N]

In one worked example, a population of 2,000 reduces the required sample from 384 to about 322 after applying the correction. This correction becomes meaningful when your calculated sample is more than about 5% of the total population.

A worked illustration 🔗

Imagine you are evaluating a new citizen grievance redressal platform in a municipality of 50,000 registered users. You want to estimate the proportion of users who are satisfied, with 95% confidence and a 5% margin of error. You have no prior data on satisfaction, so you set p = 0.5 for the most conservative estimate.

Plugging into Cochran’s formula: n₀ = (1.96)² × (0.5)(0.5) / (0.05)² = 3.8416 × 0.25 / 0.0025 = 384.16. Since 384 is less than 5% of 50,000, the finite correction barely changes anything, and your target sample is 385 users.

Now suppose you also want to compare satisfaction between two age groups. You will need enough responses in each subgroup for the subgroup comparison to be meaningful. A sample of about 800 is typically recommended for second-level breakdowns such as gender or age group comparisons. Subgroup analysis is a common reason researchers end up with samples much larger than the baseline 384.

Practical considerations often overlooked 🔗

Formulas give you a floor. In practice, a few other realities shape the final number.

Non-response adjustment 🔗

Not everyone you invite will respond. If past experience suggests a 60% response rate and you need 385 completed surveys, you should invite 385 / 0.60 ≈ 642 people. Cochran’s formula itself does not account for non-response, so researchers must inflate the calculated n by dividing by the expected response rate.

Sampling design 🔗

The standard formulas assume simple random sampling. Complex designs – stratified, clustered, or multistage – are less statistically efficient, and a design effect multiplier (often between 1.5 and 2.5) should be applied. A cluster sample of villages within districts, for instance, typically needs a larger n than a simple random sample of the same population to achieve equivalent precision.

Ethics and resources 🔗

Over-sampling wastes money, participant time, and – in clinical or behavioural studies – exposes more people than necessary to the research. The goal is not a large sample but an appropriately sized one that achieves sufficient power, with analysis techniques and results reported transparently. Undersized studies risk missing real effects; oversized ones risk finding trivial differences that look statistically significant but have no practical meaning.

Pilot studies 🔗

When the expected proportion or standard deviation is unknown, a small pilot study can sharpen your estimates before the main round. A pilot of 30 to 50 respondents often provides enough information to refine σ or p and recalculate a tighter, more efficient sample size.

Common pitfalls 🔗

Two mistakes show up repeatedly in published work. The first is treating sample size as a box to tick rather than a design decision. A sample of 100 drawn badly beats a sample of 1,000 drawn well only rarely; more often, a thoughtful 400 is stronger than a convenient 2,000. The second is forgetting that statistical significance is not the same as practical importance. A very large sample can produce p-values below the significance threshold even when the effect has no practical or clinical importance, producing false positives in terms of real-world meaning.

Reporting matters too. Whenever you publish findings, state the formula used, the values plugged in (Z, p or σ, e), any finite correction applied, the expected response rate, and the design effect. Transparency allows readers – and future researchers – to judge whether your conclusions rest on solid ground.

A quick checklist before you finalise n 🔗

Before you lock in your sample size, work through this sequence: Define the primary statistical question (mean, proportion, comparison). Choose a confidence level based on the stakes involved. Decide the acceptable margin of error by asking how precise your decision truly needs to be. Estimate variability using pilot data, prior studies, or p = 0.5 as a default. Apply the appropriate formula, add a finite correction if needed, adjust for expected non-response, and multiply by a design effect if your sampling is not simple random. The resulting number is your working target – defensible, efficient, and tied clearly to your research objective.

What do you think? If you were designing a study on citizen satisfaction with a flagship public scheme in your state, would you prioritise a tighter margin of error or a higher confidence level – and why? And how would you justify your chosen sample size to a sceptical evaluator who thinks “more is always better”?

References

1. https://www.iwh.on.ca/what-researchers-mean-by/sample-size-and-power
2. https://calculator.academy/cochrans-sample-size-calculator/
3. https://sociology.institute/research-methodologies-methods/optimal-sample-size-research-studies/
4. https://www.qualtrics.com/articles/strategy-research/determine-sample-size/
5. https://www.sopact.com/use-case/survey-sample-size-calculator
6. https://www.geopoll.com/blog/sample-size-research/
7. https://dissertationdataanalysishelp.com/cochrans-sample-size-calculator/
8. https://pmc.ncbi.nlm.nih.gov/articles/PMC10000262/
9. https://pmc.ncbi.nlm.nih.gov/articles/PMC7745163/