Key takeaways
- Probability sampling (simple random, systematic, stratified, cluster) gives every member of the population a known, non-zero chance of selection and supports statistical generalisation.
- Non-probability sampling (convenience, purposive, quota, snowball) is common and legitimate, but limits how far you can generalise.
- Simple random sampling needs a complete sampling frame and a genuinely random selection process, such as a random number generator.
- There is no single best sampling method: the right one depends on your research question, population, frame, budget and analysis.
- Qualitative studies usually use purposive sampling and aim for information-rich cases rather than statistical representativeness.
What are sampling methods?
Sampling methods are the procedures researchers use to select a subset of people, organisations, documents or events (the sample) from the larger group they want to learn about (the population). They fall into two families: probability sampling, where selection is random and each member has a known chance of inclusion, and non-probability sampling, where selection depends on access or the researcher’s judgement.
Three terms come up constantly. The target population is everyone your conclusions are meant to cover. The sampling frame is the actual list you select from, such as a student registry or a list of registered clinics. The sampling unit is what gets selected: a person, a household, a school. Gaps between the target population and the frame are a major source of bias.
Probability sampling methods, with examples
| Method | How it works | Example | Main strength / limitation |
|---|---|---|---|
| Simple random sampling | Every member of the frame has an equal chance; selection by random number generator or lottery | Randomly select 200 of 2,000 enrolled first-years for a wellbeing survey | Unbiased and easy to analyse / needs a complete list |
| Systematic sampling | Pick a random start, then every kth member, where k = N / n | From 2,000 students, k = 10: start at a random number from 1 to 10 (say 7), then take 7, 17, 27… | Simple to carry out / biased if the list has a pattern repeating every k entries |
| Stratified sampling | Divide the population into non-overlapping subgroups (strata), then randomly sample within each | Sample separately from each faculty so every faculty is represented in proportion | More precise estimates; guarantees subgroup coverage / needs stratum information in the frame |
| Cluster sampling | Randomly select whole groups (clusters), then include everyone or a random subsample within them | Randomly choose 15 of 120 secondary schools, then survey all Year 10 pupils in those schools | Practical for dispersed populations / less precise, because people within a cluster tend to be similar |
| Multistage sampling | Combine stages, such as regions, then schools, then classes | National education surveys | Feasible at scale / complex weighting and analysis |
Stratified and cluster sampling are often confused. In stratified sampling you sample from every group; in cluster sampling you sample some groups and study them whole. Cluster designs need analysis methods that account for clustering, such as multilevel models, or the standard errors and p-values will be too optimistic.
How to do simple random sampling, step by step
Simple random sampling is the benchmark every other method is compared with. Here is how to do it properly.
- Define the population precisely. “Full-time undergraduates enrolled in the 2025–26 academic year at one university”, not just “students”.
- Obtain a complete sampling frame. For example, a registry list of 2,000 eligible students. Remove duplicates and ineligible entries.
- Number every unit from 1 to N.
- Decide the sample size (see below), say n = 200.
- Generate random numbers. Use software, for example
sample(1:2000, 200)in R or RAND() in Excel sorted in order, and keep a record of the seed so the draw can be reproduced. - Contact everyone selected and track who responds. Replacing non-responders with whoever is convenient quietly turns a random sample into a convenience sample.
Sampling “without replacement” (each person can be chosen only once) is standard in research. Note that random sampling (who is in the study) is different from random assignment (who gets which condition in an experiment). The first supports generalisation; the second supports causal conclusions.
Non-probability sampling methods
- Convenience sampling: whoever is easiest to reach, such as students in your own seminar. Fast and cheap, but the most open to bias. Be candid about it in your limitations.
- Purposive (judgement) sampling: deliberately selecting cases that suit the research question, such as head teachers who have led a school out of special measures. The standard approach in qualitative research.
- Quota sampling: setting targets for subgroups (for example 50% women, 30% part-time) and filling them non-randomly. It looks like stratified sampling but lacks random selection within groups.
- Snowball sampling: participants refer others. Useful for hidden or hard-to-reach populations, such as undocumented workers or people with a rare condition, but networks tend to recruit similar people.
- Volunteer (self-selection) sampling: people respond to an advert or online link. Very common in online surveys; respondents are often more interested in the topic than non-respondents.
Non-probability samples are not wrong. They are often the only realistic option for a master’s dissertation. The key is to describe the method honestly, explain who might be missing, and avoid claiming statistical generalisation to a population you did not sample from.
Want a second pair of eyes? Designing a study and unsure how to sample or how many participants you need? We can review your plan with you and explain the options. See methodology chapter help.
Sampling methods for qualitative research
Qualitative studies aim for depth and insight rather than estimates, so they rarely use random sampling. Common strategies include:
- Maximum variation: deliberately including very different cases to see what patterns hold across them.
- Homogeneous sampling: similar participants, to explore one experience in depth.
- Typical and extreme (deviant) case sampling: the ordinary, or the unusually successful or unsuccessful.
- Criterion sampling: everyone meeting a defined criterion, such as nurses who left the profession within two years.
- Theoretical sampling: in grounded theory, choosing later participants based on concepts emerging from earlier interviews.
Sample size in qualitative work is usually justified by information power or saturation (the point where new data add little new insight), not by a formula. State your reasoning in your methodology chapter rather than quoting a magic number.
How large should your sample be?
For quantitative studies, sample size should follow from your analysis. Two common approaches:
Estimating a proportion. To estimate a percentage within a chosen margin of error e:
n₀ = z² · p(1 − p) / e²
With 95% confidence (z = 1.96), no prior estimate (p = 0.5, the most conservative choice) and e = 0.05: n₀ = 3.8416 × 0.25 / 0.0025 = 384.16, so 385 respondents. For a small population, apply the finite population correction n = n₀ / [1 + (n₀ − 1) / N]. With N = 2,000: 384.16 / (1 + 383.16 / 2,000) ≈ 322.4, so 323 respondents. Then inflate for expected non-response: at a 60% response rate, invite about 323 / 0.6 ≈ 539 people.
Testing a hypothesis. Use a power analysis (for example in G*Power or R) based on the expected effect size, alpha and desired power. For instance, an independent-samples t-test needs about 64 per group to detect a medium effect (d = 0.5) with 80% power at α = .05, two-tailed. Cluster designs need more again, to offset the similarity within clusters.
Sampling error and sampling bias
Sampling error is the natural difference between a sample statistic and the population value that arises because you studied only part of the population. It is random, shrinks as the sample grows, and is exactly what standard errors and confidence intervals quantify.
Sampling bias is systematic: some members of the population are more likely to be included than others in ways related to what you are measuring. A bigger sample does not fix it. The main forms are:
- Coverage bias: the frame misses part of the population, such as an online-only survey of older adults.
- Non-response bias: those who reply differ from those who do not. Compare early and late responders, or responders with known population figures, to gauge it.
- Self-selection bias: volunteers are typically more engaged with the topic.
- Survivorship bias: studying only cases that remain, such as firms still trading or students who completed the course.
Report your response rate, describe how your achieved sample compares with the population on key characteristics, and discuss any likely direction of bias in your limitations.
How to choose the best sampling method
Work through these questions:
- Do you need to generalise statistically to a defined population? If yes, use a probability method. If your aim is exploration or theory-building, purposive sampling may fit better.
- Do you have a complete sampling frame? If yes, simple random or systematic sampling is feasible. If not, consider cluster or multistage sampling.
- Are subgroup comparisons central? Stratify, and consider oversampling small groups (then weight in analysis).
- Is the population large and spread out? Cluster sampling usually saves time and cost.
- Is the population hidden or hard to reach? Snowball or respondent-driven sampling.
For large populations with a good frame and a quantitative design, stratified random sampling is often the strongest practical choice. For correlational research, any method that captures the full range of both variables helps, since a restricted range weakens correlations. Whatever you choose, justify it in your methodology chapter and link it to your limitations. If you want feedback on that justification, our methodology chapter help and survey and questionnaire design support can help you sharpen it.