Sampling methods are the ways researchers choose which people, places or items to study from a larger population. They fall into two groups: probability sampling, where every member of the population has a known chance of selection (simple random, systematic, stratified and cluster sampling), and non-probability sampling, where the researcher’s choice or participants’ availability decides who takes part (convenience, quota, purposive and snowball sampling).
The difference matters because it decides what you can claim. With a probability sample you can estimate how far your results might differ from the whole population. With a non-probability sample, as the UK public health textbook HealthKnowledge puts it, ‘you cannot estimate the effect of sampling error’.
If you’re planning the data collection for a dissertation or thesis, this guide compares the eight main methods in one table, works through an example of each, looks at how two real UK studies chose their samples, and ends with a template for justifying your choice in your methodology chapter.
Key terms
- Population: everyone (or everything) your research question is about, for example all postgraduate researchers at UK universities.
- Sampling frame: the list you actually draw the sample from, such as a university’s register of enrolled students. Probability sampling needs one: HealthKnowledge describes it as starting ‘with a complete sampling frame of all eligible individuals’.
- Sample: the people or items you select and study.
- Sampling bias: a sample that differs from the population in a way that distorts the results. HealthKnowledge lists low response rates, an out-of-date sampling frame and leaving out hard-to-reach people among its causes.
The eight sampling methods compared
| Method | How it works | When it fits | Main risk |
|---|---|---|---|
| Simple random (probability) | Each member of the sampling frame has an equal chance of selection, for example by random number generator | You have a complete list and a fairly uniform population | May pick too few people from a small subgroup |
| Systematic (probability) | Pick every kth member of the list after a random start | Long lists where drawing random numbers is awkward | Bias if the list has a repeating pattern |
| Stratified (probability) | Divide the population into subgroups (strata) that share a characteristic, then sample randomly within each | Subgroups you must compare or represent fairly | Needs the characteristic recorded on the sampling frame |
| Cluster (probability) | Randomly select whole groups (hospitals, wards, postcode areas), then study everyone or a sample within them | Populations spread over a wide area, or with no list of individuals | Bias if the chosen clusters aren’t representative |
| Convenience (non-probability) | Recruit whoever is available and willing | Pilot studies and early exploratory work | Results prone to significant bias |
| Quota (non-probability) | Recruit set numbers of people with given characteristics (for example 50 men and 50 women) until each quota is full | Quick surveys that need a rough match to the population on a few characteristics | Unrepresentative on characteristics the quotas didn’t cover |
| Purposive (non-probability) | The researcher chooses participants who fit the purpose of the study | Qualitative research needing people with particular experience or knowledge | Errors of judgement; not representative |
| Snowball (non-probability) | Participants nominate others they know | Hidden or hard-to-reach groups with no sampling frame | Significant risk of selection bias |
The definitions and risks follow HealthKnowledge and an educational review of sampling in clinical research by Zrineh, Al-Usta and Alwawi (2026).
A worked example of each method
To compare the methods on one problem, imagine a master’s dissertation on how postgraduate researchers at one UK university feel about their supervision. The example is invented for this guide. Suppose the graduate school can give you an anonymised list of its 2,000 enrolled postgraduate researchers, each tagged with their faculty, and you want 200 survey responses.
Probability sampling
- Simple random sampling: number the list from 1 to 2,000 and use a random number generator to pick 200 numbers. Every researcher has the same 1-in-10 chance of selection.
- Systematic sampling: with 2,000 names and 200 places, take every 10th name after a random start between 1 and 10. Check first that the list isn’t ordered in a way that repeats every 10 names (for example, by supervisor group in blocks of ten), or you could select only one kind of student.
- Stratified sampling: say 40% of the researchers are in science and engineering, 30% in medicine and health, and 30% in arts, humanities and social sciences. Sample randomly within each faculty in the same proportions: 80, 60 and 60. Each faculty is now represented in its true share. If one faculty is small but you need to compare it with the others, you can sample it more heavily and weight the results in your analysis.
- Cluster sampling: if you had no central list and could only reach students through their departments, you could randomly select some departments and survey everyone in them. It’s practical, but students in the same department share supervisors and culture, so if the departments you happen to select aren’t typical, neither are your results. Selecting clusters and then sampling within them is multistage sampling, explained step by step in our guide to multistage sampling.
Non-probability sampling
- Convenience sampling: you post the survey link in your own department’s group chat. Responses arrive quickly, but they come from people you know, in one department, who were online that week.
- Quota sampling: you ask for responses until you have 80 from science and engineering, 60 from medicine and health and 60 from the other faculties. The faculty mix now matches the population, but within each faculty you still have whoever responded first.
- Purposive sampling: for follow-up interviews you choose 12 researchers who changed supervisor during their PhD, because that experience is what your research question is about.
- Snowball sampling: to reach researchers who left their PhD without finishing (there’s no list of them), you ask each person you interview to pass your details to others in the same position.
Two real UK studies and how they sampled
Scottish Health Survey: stratified, clustered and systematic in one design
The Scottish Health Survey 2022 technical report shows how a large national survey combines probability methods:
- Sampling frame: Royal Mail’s small user Postcode Address File, a list of postal addresses.
- Stratification: the main sample was divided into 32 strata, one for each local authority, so every council area is represented.
- Clustering: a two-stage clustered design, with small areas (intermediate geographies) randomly selected at the first stage.
- Systematic selection: within each selected area, addresses were chosen by systematic random sampling, after ordering them by urban-rural classification, deprivation rank and postcode.
The lesson for your own project: methods can be combined, and your methodology chapter should say what was done at each stage and why.
UK Biobank: a volunteer sample and healthy volunteer bias
The UK Biobank health study invited people aged 40 to 69 living near its assessment centres, and those who chose to take part formed its sample. A 2017 study by Fry and colleagues in the American Journal of Epidemiology reports that about 9.2 million people were invited and 5.5% took part.
Those who joined were not a random cross-section. Compared with non-participants, they were older, more often female and from less deprived areas; compared with the general population, they were less likely to be obese, to smoke or to drink daily. The authors conclude that the cohort is not representative of the population it was drawn from, with evidence of a ‘healthy volunteer’ selection bias. They also argue that relationships between exposures and disease found in the cohort may still be widely generalisable.
The lesson: a non-probability sample can still answer some questions well, but you must say which questions it can’t answer, and why.
Probability vs non-probability sampling: which should you use?
| Probability sampling | Non-probability sampling | |
|---|---|---|
| Needs a sampling frame? | Yes | No |
| Can you estimate sampling error? | Yes | No |
| Can you generalise statistically to the population? | Yes, if response is good | No: the Zrineh review says purposive sampling ‘limits the ability to make statistical inferences to broader populations’ |
| Cost and time | Higher | Lower |
| Fits | Surveys and experiments that estimate how common something is, or test hypotheses | Qualitative studies, pilot work, hard-to-reach groups |
Four questions decide it:
- What will you claim? If your conclusion is a figure for the population (‘35% of postgraduate researchers…’), you need a probability sample. If you want to understand experiences in depth, purposive sampling fits better.
- Do you have a sampling frame? No list means no simple random, systematic or stratified sample. Cluster sampling, or a non-probability method, may be the realistic choice.
- Do you need to compare subgroups? Stratify, so each subgroup has enough people.
- What can you afford? Time, travel and access are legitimate reasons, as long as you state the limitations they create.
How big should your sample be?
For quantitative studies that test a hypothesis, sample size can be set by a power calculation: the number of participants needed to detect the effect you expect. Statistical power is the probability of rejecting a false null hypothesis, and the Zrineh review gives 80% or 90% as conventional levels; higher power needs a larger sample. Your supervisor or a statistician can help you set the expected effect size. Our guide to the statistical treatment of data covers the analysis that follows.
For a qualitative study, explain your number in terms of the depth of data your analysis needs. Whatever the method, say how you decided the number.
Common sampling mistakes in dissertations
- Calling a convenience sample random. Sending a survey to everyone you can reach isn’t random sampling. Name the method you actually used.
- Generalising from a non-probability sample. Twelve purposively chosen interviewees can show how something is experienced, not how common it is.
- Ignoring non-response. A random sample in which few people reply can still be biased: low response rates are one of the causes of sampling bias listed above. Report your response rate and who didn’t respond.
- Confusing stratified and cluster sampling. In stratified sampling you sample from every group; in cluster sampling you sample some groups and study everyone or a sample within them.
- Mixing up the unit of analysis. If you sample departments but analyse individual students, say so. Our guide to the unit of analysis explains why it matters.
How to justify your sampling method in your methodology chapter
Your examiners will want to know the population, the sampling frame, the method, the size and the limitations. A template you can adapt:
‘The target population was [population]. Participants were drawn from [sampling frame], which listed [number] eligible [people/items]. A [method] sample was used because [reason linked to the research question], [with strata defined by X / with clusters defined by Y]. The sample size of [n] was set by [a power calculation for an effect size of Z at 80% power / the depth of data needed for the qualitative analysis / practical constraints]. [Number] [people] took part, a response rate of [x]%. Because [limitation, for example respondents were self-selected], the findings [can / cannot] be generalised to [population], and this is discussed in [section].’
Your methods chapter should also describe the tools you used to collect data; see our guide to research instruments. For where your study design sits among quantitative, qualitative and mixed approaches, see types of research, and to set out what your sample does and doesn’t cover, see scope and delimitations.
Next steps
Write down your population, your sampling frame and what you want to claim before choosing a method; the four questions above then narrow the choice. If you’re thinking about a research degree where you’d design studies like these, you can search current PhD opportunities.
