Sample Size

Sample size is the number of participants, cases or observations included in a study. In a dissertation, choosing it is less important than justifying it: the number must follow from what the study is trying to do, whether that is estimating a population value precisely, detecting an effect with adequate statistical power, or gathering enough qualitative depth to answer the research question.

For a dissertation, the practical question is not “what is the right sample size?” There is no single right number. The question is:

What is my study trying to achieve, which method of determining sample size fits that aim, and how will I explain the number to my examiner?

This page separates the three logics that are most often confused (precision, statistical power and qualitative adequacy), works through the calculations a student is most likely to need, and shows how to write the justification in a methodology chapter. Detailed guidance on qualitative stopping decisions is on the separate page on data saturation.

On this page:

Sample Size Explained Simply

Imagine three students at the same business school. The first wants to estimate what percentage of the school’s 1,200 undergraduates use AI tools for coursework. The second wants to test whether students who receive a short training session score higher on a referencing test than students who do not. The third wants to understand how a small number of lecturers have changed their assessment practices because of AI.

All three must answer “how many participants?”, but they need different kinds of answer. The first needs enough respondents for the percentage to be estimated within an acceptable margin of error. The second needs enough participants to have a good chance of detecting a training effect if one of a meaningful size exists. The third needs enough lecturers, and enough depth in each interview, to understand the changes well. A sample size that is generous for one aim can be inadequate for another, which is why the aim has to come first.

Dudovskiy Research Assistant

Not sure how many participants your study needs?

Enter your topic to receive a free, tailored methodology preview.

Free · No account required
Learn more about Dudovskiy Research Assistant

Six Ways to Justify a Sample Size

Lakens (2022) sets out six ways in which researchers justify the sample size of a quantitative study. They are useful for dissertations because they make clear that a calculation is only one of several honest options, and that “no justification” is also a choice that should be stated openly.

Justification What it means When it fits a dissertation
Measure the entire population The population is finite, can be listed, and (almost) every member can be included A small, defined population, such as all 40 managers of one firm
Resource constraints Time, money or access to participants set the number Common in student projects; should be paired with an honest statement of what the sample can and cannot show
Accuracy Enough data to estimate a value with a desired precision Descriptive surveys estimating percentages or averages
A-priori power analysis Enough data to detect an effect of a specified size with a desired probability Experiments, group comparisons, correlation and regression studies testing hypotheses
Heuristics A rule of thumb or norm from the literature Weak on its own, unless the reasoning behind the rule fits your study
No justification No specific reason or inferential goal, stated openly Rarely acceptable in a dissertation, but more honest than an invented justification

Lakens (2022) notes that resources always limit sample size, so resource constraints are usually at least a secondary justification even when another approach is the main one. He is critical of heuristics used without understanding: a rule of thumb is often trusted because of who recommended it rather than because it fits the study. For qualitative studies, which these six categories were not designed for, the logic is different and is covered below.

The figure below is a guide to choosing between the main approaches. It is a guide, not a rule: a study with several aims may need more than one route.

Decision diagram: depending on whether a study estimates a value, tests a hypothesis, has a sample fixed by access, studies a small listable population or is qualitative, use precision, power analysis, sensitivity analysis, a census or sample adequacy, with what to report for each
Figure: Choosing how to determine your sample size. A guide, not a rule; the author’s synthesis based on Lakens (2022), Cochran (1977) and Malterud et al. (2016).

Sample Size for Estimating a Population Value

When the aim is to estimate a value, such as the percentage of customers who would recommend a brand, the sample size depends on how precise the estimate needs to be. Precision is usually expressed as a margin of error at a stated confidence level. This logic assumes a probability sample; with a non-random sample, a margin of error cannot be interpreted in the usual way.

For a proportion, the standard formula for a large population (Cochran, 1977) is:

n₀ = z² × p(1 − p) / e²

where z is the value for the chosen confidence level (1.96 for 95%), p is the expected proportion and e is the margin of error expressed as a decimal. When the expected proportion is unknown, p = 0.5 is used because it gives the largest, most cautious sample size.

Worked example. For a 95% confidence level and a margin of error of ±5 percentage points: n₀ = 1.96² × 0.5 × 0.5 / 0.05² = 3.8416 × 0.25 / 0.0025 = 384.16, which is rounded up to 385 completed responses.

If the population is small and known, a finite population correction reduces the number: n = n₀ / (1 + (n₀ − 1) / N). For the 1,200 undergraduates in the earlier example, n = 384.16 / (1 + 383.16 / 1,200) = 291.2, rounded up to 292. The table below shows how the required number changes with population size, using the same assumptions.

Population size (N) Required completed responses (95%, ±5 points, p = 0.5)
300 169
500 218
1,200 292
5,000 357
100,000 383
Very large 385

Two points follow. First, beyond a few thousand, population size makes little difference: precision depends mainly on the number of responses, not on the proportion of the population sampled. Second, the formula gives completed responses, not invitations. If you expect 40% of those invited to respond, you need to invite about 730 people to obtain 292 responses (292 / 0.40 = 730). Inviting more people corrects the number of responses, but it does not correct non-response bias; people who reply may differ from those who do not (see sampling bias).

The precision you need also depends on how the results will be used. With 100 responses, the margin of error for a proportion near 50% is about ±10 percentage points; with 385, it is about ±5. If you plan to report results separately for subgroups, such as departments or customer segments, each subgroup needs enough responses for its own estimate, which is often the real driver of total sample size.

About the Krejcie and Morgan table. Many dissertations cite Krejcie and Morgan (1970), whose table gives sample sizes for populations of different sizes. The table is calculated from a formula for estimating a proportion with p = 0.5, a margin of error of 0.05 and a 95% confidence level, which is why its figures match the table above to within one or two cases (291 for a population of 1,200, for example) and level off at slightly more than 380. Citing it is reasonable only when those assumptions match your study: a random sample, an estimate of a proportion and that level of precision. It does not determine the sample size needed to test a hypothesis, to run a regression or factor analysis, or to conduct a qualitative study, and it cannot make a convenience sample representative.

Sample Size for Testing a Hypothesis: Power Analysis

When the aim is to test a hypothesis, for example that two groups differ or that two variables are related, the sample size is usually determined by an a-priori power analysis carried out before data collection. It depends on four quantities, and fixing any three determines the fourth:

Quantity What it means Usual choice in a dissertation
Significance level (alpha) The accepted risk of concluding there is an effect when there is none (Type I error) 0.05
Statistical power The probability of detecting an effect of the specified size if it exists (1 minus the Type II error rate) 0.80 is common; 0.90 is stricter
Effect size The size of the difference or relationship the study should be able to detect The hardest choice; it must be justified (see below)
Sample size The number of participants needed The result of the calculation

The effect size is the decision that most needs justifying. Lakens (2022) regards a smallest effect size of interest, the smallest effect that would matter theoretically or practically, as the strongest basis for a power analysis. Where that cannot be specified, an estimate from previous studies can be used, with caution, because effects reported by small studies tend to be inflated (Button et al., 2013). Cohen (1988) proposed conventions for small, medium and large effects (for a difference between two means, d = 0.2, 0.5 and 0.8), but he also noted that these labels are relative to the field and the context. Lakens (2022) argues that these benchmarks should not be used as the basis of an a-priori power analysis, even though “medium” is a common default in software.

Worked example. The table below shows the number of participants needed per group to compare two independent groups with a two-sided t-test, alpha = 0.05 and power = 0.80. The figures were calculated with the noncentral t distribution and match those produced by standard power analysis software.

Effect size (Cohen’s d) Participants per group Total sample
0.8 26 52
0.5 64 128
0.3 176 352
0.2 394 788

The table shows why the effect size matters so much: halving the effect you want to detect roughly quadruples the sample you need. Raising power from 0.80 to 0.90 for d = 0.5 increases the requirement from 64 to 86 per group. For a correlation, about 85 participants are needed to detect r = 0.30 with the same alpha and power, using the common Fisher z approximation (exact software routines may give one fewer).

The consequences of ignoring power are practical as well as statistical. A study with 30 participants per group has only about a 48% chance of detecting a true medium effect (d = 0.5), so a non-significant result says little. Button et al. (2013) showed that low power not only reduces the chance of detecting true effects but also reduces the likelihood that a statistically significant result reflects a true effect, and tends to inflate the effect sizes that are reported.

When the sample is fixed. If access, not calculation, sets your sample size, a sensitivity power analysis is the honest alternative: instead of asking how many participants you need, you report the smallest effect your sample could detect with adequate power. For example, with 50 participants per group, alpha = 0.05 and power = 0.80, a two-group comparison can reliably detect only effects of about d = 0.57 or larger. Stating this lets the examiner judge what your non-significant results can and cannot mean.

Free software such as G*Power covers power analysis for the t, F and chi-square families of tests, including ANOVA, regression and the chi-square test (Faul et al., 2007). Whatever tool you use, report the test, the inputs (alpha, power, effect size and its source) and the result, so that the calculation can be repeated.

Sample Size in Qualitative Research

Qualitative studies do not use precision or power calculations, because they do not aim to estimate population values or test statistical hypotheses. Sample size is judged by whether the data are sufficient to answer the research question in depth. The two rationales most often used are saturation, the point at which further data add little new information, and information power, which links the sample size to the aim, the specificity of the participants, the use of theory, the quality of the dialogue and the analysis strategy (Malterud et al., 2016).

Published studies give useful planning ranges, but no universal number. Guest et al. (2006) found that saturation was reached within the first 12 of 60 interviews in a fairly homogeneous group. A systematic review of empirical tests found saturation within 9 to 17 interviews or 4 to 8 focus groups, mainly in studies with relatively homogeneous populations and narrowly defined objectives (Hennink and Kaiser, 2022). Broader questions and more varied participants need more. These figures describe when saturation was observed under particular conditions; they are not targets to quote as proof that your own sample was sufficient. The same caution applies to the typical ranges given on the non-probability sampling page: they are starting points for planning, not justifications.

For how to decide when to stop, how to document it and why saturation does not suit every qualitative approach, see data saturation.

What to Write in Your Methodology Chapter

Whatever the approach, a convincing justification states the aim, the method used to determine the number, the inputs and their sources, the result, and how the achieved sample compares with the plan. The table below summarises what to write for the most common dissertation designs.

Study type How to determine the sample size What to report
Descriptive survey with a probability sample Precision formula with finite population correction, then adjust for expected response rate Confidence level, margin of error, p used, population size, expected and achieved response rates
Survey with a non-probability sample Resource constraints or the needs of the planned analysis; no margin of error Why the number was chosen, the sampling method and the limits on generalisation
Experiment or group comparison A-priori power analysis Test, alpha, power, effect size and where it came from, software, required and achieved sample
Correlation or regression study A-priori power analysis for the planned test As above, including the number of predictors in a regression
Fixed or opportunity sample Sensitivity power analysis The smallest effect that could be detected with the achieved sample
Small defined population Census of the whole population Population size, number reached and reasons for any non-participation
Qualitative interview study Information power and, where it fits the approach, saturation Planned range and its rationale, stopping decision and how it was documented

Two habits strengthen any justification. Decide and record the sample size before collecting data, so that the number cannot be adjusted to the results. And report the achieved sample honestly: if you planned 128 participants and obtained 90, say so and explain what that means for power or precision, rather than leaving the reader to work it out.

Sample Size in Business Research

Business dissertations often study populations that are small and hard to reach, such as the senior managers of one organisation, the small firms in one town or the suppliers of one company. When the whole population can be listed and is small, a census is often more defensible than a sample, and the justification becomes a matter of how many members were reached and why some were not.

Organisational surveys also tend to face low response rates, because managers and employees receive many requests. Planning for this means estimating a realistic response rate, inviting enough people, and recognising that a large number of responses from a self-selected group is still a self-selected group. When responses are fewer than planned, report a sensitivity analysis or a widened margin of error rather than quietly treating the smaller sample as adequate. Where the sample comes from the student’s own network, the justification is usually a resource constraint, and the conclusions should be framed accordingly.

Application Example

Hypothetical example. A master’s student plans an online experiment testing whether showing a sustainability label increases customers’ willingness to pay for coffee. Participants will be randomly assigned to see the product with or without the label.

She first looks for a smallest effect size of interest. A café owner she interviewed during the proposal stage says that an increase of less than about 20 pence would not justify the cost of certification. Using the standard deviation of willingness to pay from a pilot of 30 people (about 40 pence), she converts this to d = 0.5. An a-priori power analysis for a two-sided independent t-test with alpha = 0.05 and power = 0.80 gives 64 participants per group. Expecting about 10% of responses to fail an attention check, she needs at least 143 participants (128 / 0.90 = 142.2, rounded up) and plans to recruit 145 to allow a small margin.

She obtains 120 usable responses. In her methodology chapter she reports the planned calculation, the achieved sample and a sensitivity analysis showing that 60 per group can detect effects of about d = 0.52 with 80% power, which is close to her smallest effect of interest. In the discussion, she notes that the pilot’s standard deviation was based on only 30 people, so the conversion from pence to d is approximate.

Common Mistakes

  • Choosing a number first and searching for a justification afterwards.
  • Citing the Krejcie and Morgan table for an experiment, a regression model or a qualitative study.
  • Reporting a margin of error for a convenience or snowball sample.
  • Calculating the number of completed responses needed but inviting only that number of people.
  • Using Cohen’s “medium” effect size as a default without explaining why it is relevant.
  • Quoting a rule of thumb, such as a minimum of 30, as if it were a justification.
  • Treating a large sample as protection against bias.
  • Ignoring subgroups that will be analysed separately.
  • Interpreting a non-significant result from an underpowered study as evidence of no effect.
  • Quoting a published number of interviews as proof that a qualitative sample was sufficient.

Sample Size in the Age of AI and Digital Research

Online panels, social media recruitment and web scraping have made large samples cheap. A survey shared online can attract thousands of responses, and a scraping tool can collect hundreds of thousands of reviews or posts. This changes the sample size problem rather than solving it. When data are plentiful, precision and power are rarely the limiting factors; the question becomes who the data represent and whether the responses are genuine. A very large sample from a self-selected online audience still describes that audience, and the narrow confidence intervals that come with large numbers can make biased estimates look convincingly precise.

Data quality is a related concern. Online surveys can attract inattentive, duplicate or automated responses, and screening them out reduces the usable sample. A sample size plan for an online study should therefore state in advance how responses will be screened, for example with attention checks or completion-time rules, and how many exclusions are expected, so that the target number refers to usable responses.

Large language models can generate synthetic survey answers or interview transcripts. These cannot count towards the sample size of a study of real people, because they do not provide information about real participants. If AI-generated responses are used at all, for example to pilot a questionnaire, they should be reported separately and excluded from the analysed sample.

AI assistants and online calculators can also perform sample size and power calculations. They are convenient, but the answer is only as sound as the inputs, and a chatbot may apply the wrong formula or test without saying so. Check any AI-produced figure against dedicated software or the formulas on this page, and report the inputs and the tool used, as you would for any other calculation.

Dissertation Example

The following paragraphs show how sample size can be justified in a methodology chapter for two hypothetical studies.

Descriptive survey. “The population consisted of 1,200 undergraduate students enrolled at the business school in the 2026/27 academic year, and a simple random sample was drawn from the school’s enrolment list. The required sample size was calculated using Cochran’s (1977) formula for estimating a proportion, with a 95% confidence level, a margin of error of ±5 percentage points and p = 0.5, because no reliable prior estimate was available. With the finite population correction, the required sample was 292 completed responses. Assuming a response rate of 40%, based on previous surveys at the school, 730 students were invited. A total of 251 usable responses were received, giving a margin of error of approximately ±5.5 percentage points for proportions near 50% (with the finite population correction), which is acknowledged in the limitations section.”

Experiment. “An a-priori power analysis was conducted in G*Power (Faul et al., 2007) for a two-sided independent samples t-test with alpha = 0.05 and power = 0.80. The smallest effect size of interest was set at d = 0.5, derived from a practically meaningful difference of 20 pence in willingness to pay and the standard deviation observed in a pilot study (n = 30). The analysis indicated a required sample of 64 participants per group (128 in total). To allow for an expected 10% exclusion rate from failed attention checks, 145 participants were recruited.”

Both paragraphs state the aim, the method, the inputs and their sources, the result and the achieved sample. The first also reports the precision actually achieved, which an examiner can check.

Exam Tip

If you are asked to evaluate a study’s sample size, start by identifying what the study was trying to do. A sample can only be judged against its aim: estimating a value calls for precision, testing a hypothesis calls for power, and a qualitative study calls for depth and relevance. Then check whether the method of justification matches the aim, whether the inputs were explained and whether the achieved sample was compared with the plan. A small sample with an honest, well-reasoned justification can be stronger than a large one that rests on a misapplied table or an unexplained rule of thumb.

Not sure how many participants your study needs or how to justify the number? Dudovskiy Research Assistant can review your topic, research questions and planned design, suggest which approach to sample size fits your study and help you word the justification in your methodology chapter. It supports your own judgement; the final decisions remain yours and your supervisor’s.

John Dudovskiy

References

Button, K.S., Ioannidis, J.P.A., Mokrysz, C., Nosek, B.A., Flint, J., Robinson, E.S.J. and Munafò, M.R. (2013). Power failure: why small sample size undermines the reliability of neuroscience. Nature Reviews Neuroscience, 14(5), 365–376. https://doi.org/10.1038/nrn3475

Cochran, W.G. (1977). Sampling Techniques (3rd ed.). New York: John Wiley & Sons.

Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Hillsdale, NJ: Lawrence Erlbaum Associates.

Faul, F., Erdfelder, E., Lang, A.-G. and Buchner, A. (2007). G*Power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior Research Methods, 39(2), 175–191. https://doi.org/10.3758/BF03193146

Guest, G., Bunce, A. and Johnson, L. (2006). How many interviews are enough? An experiment with data saturation and variability. Field Methods, 18(1), 59–82. https://doi.org/10.1177/1525822X05279903

Hennink, M. and Kaiser, B.N. (2022). Sample sizes for saturation in qualitative research: a systematic review of empirical tests. Social Science & Medicine, 292, 114523. https://doi.org/10.1016/j.socscimed.2021.114523

Krejcie, R.V. and Morgan, D.W. (1970). Determining sample size for research activities. Educational and Psychological Measurement, 30(3), 607–610. https://doi.org/10.1177/001316447003000308

Lakens, D. (2022). Sample size justification. Collabra: Psychology, 8(1), 33267. https://doi.org/10.1525/collabra.33267

Malterud, K., Siersma, V.D. and Guassora, A.D. (2016). Sample size in qualitative interview studies: guided by information power. Qualitative Health Research, 26(13), 1753–1760. https://doi.org/10.1177/1049732315617444

Last reviewed: November 2026.

[]