Sampling methods: which one fits your survey?

Compare random, stratified, cluster, quota and convenience sampling. Learn how to choose a method, draw a sample and know what your results can support.

The short answer: in a random sample, every person in the population has a known, nonzero probability of being selected. Random selection supports statistically grounded conclusions about the population, provided you account for gaps in the sampling frame and nonresponse, and use an appropriate method of analysis (Jacob et al. 2019, pp. 73-74). Quota sampling, purposive sampling, convenience sampling and snowball sampling are often quicker and cheaper, but their results initially describe only the respondents. Generalizing beyond them requires assumptions that you must state explicitly.

So the method you choose depends less on your budget than on what you want to claim at the end. In this article I'll answer five questions in turn:

1. What are the population, the sampling frame and the sample?
2. What are the main random sampling methods, and when should you use each one?
3. How do you calculate a stratified sample with weighting?
4. What can quota, convenience and snowball samples and online panels do, and what can't they?
5. Which method fits a thesis, an employee, customer or population survey?

At a thesis defense, the question almost always comes up: "How did you draw your sample?" In companies, it's management who asks, usually in the form "And that applies to everyone?" The honest answer is often "I posted the link in three groups" or "We put it on our website." That's not a disaster. It's a sampling method with a name, with strengths and with clear limits. It only gets awkward when the report still says that "students across the country" or "our customers" think this or that. To keep that from happening to you, we'll work through the methods and check which claims each one supports. How many responses you need is not the topic here. That's covered in the article Calculating sample size and in the sample size calculator.

Population, sampling frame, sample

Before you choose a method, you need three terms. They sound dry, but they decide almost everything that can go wrong later.

The population includes everyone you want to say something about: all employees of your company, all customers of the last year, all students at your university. Schumann (2019, p. 84) calls it the intended population or target population. The more precisely you define it, the easier everything else becomes. "Our customers" is too vague. "Everyone who ordered from our online store at least once in 2026" is usable.

The sampling frame is the list you actually draw from: the staff list, the customer database, the university's email directory. Malhotra and Birks (2007, p. 407) describe it as a representation of the target population, either a list or a set of directions for finding its members. The list almost never matches the population exactly. Some people are missing, some are on it although they don't belong, some appear twice. Anyone who isn't on the list has no chance of being drawn. Jacob et al. (2019, p. 75) illustrate this with the German residents' register, where people are required to record their address: students who live in a city but haven't registered there are missing from every sample drawn from that register.

The sample is the set of people you select from the sampling frame. Only some of them will respond. So whether you can generalize your results to everyone depends not only on the draw, but also on who is missing from the list and who doesn't respond. Schumann (2019, pp. 84-85) distinguishes the intended population, the frame population and the inference population, and advises making sure they coincide as closely as possible before the survey, not only when interpreting the results.

By the way, a census has these three levels too. If you invite all 180 employees, the staff list is your sampling frame, and anyone who doesn't respond is still missing. You no longer need a selection method, but the question of who is missing remains. More on this in the article on survey response rate.

Random sampling: five methods explained

You have a random sample (also called a probability sample) when every person in the population has an equal or at least calculable chance greater than zero of getting into the sample. This requires the population to be known and precisely defined, and every person may appear only once (Jacob et al. 2019, p. 74). So "random" doesn't mean "somehow." If you approach a hundred passers-by in the town square, you're not drawing a random sample, even if it feels like one (Jacob et al. 2019, p. 86).

Only with random samples can the uncertainty caused by selection be determined statistically, for example with a confidence interval. The calculation has to match the sampling method, and it doesn't automatically capture gaps in the list or nonresponse that distorts the results. With all other methods, every generalization needs additional assumptions that you disclose (Jacob et al. 2019, pp. 73-74, and Malhotra and Birks 2007, p. 411).

Simple random sampling

Think of a lottery drum. Every person has the same chance, and every combination of people is equally likely (Malhotra and Birks 2007, p. 415). In practice, you number your list and let a program generate random numbers. In Excel, for example: add a column with =RAND() next to the list, sort by that column and take the first n rows.

Simple random sampling is easy to explain, and most statistical methods assume it. Its drawback: small groups may be underrepresented or missing entirely by chance, especially in small samples (Malhotra and Birks 2007, p. 416). If you draw from a workforce with only a few salespeople, sales may be almost absent from your sample. Stratification helps here.

Systematic random sampling

Here you choose a random starting point and then take every i-th person on the list. The interval i is the size of the population divided by the desired sample size. With 100,000 addresses and a target sample size of 1,000, i = 100. If the random start is 23, you draw numbers 23, 123, 223 and so on (Malhotra and Birks 2007, p. 416).

The method is simpler than simple random sampling and even works without a list, for example if you survey every tenth person leaving a store. But check the order of your list. If it's sorted by revenue, the sample covers small and large customers evenly. If it follows a cycle, such as monthly data with an interval of 12, you'll always draw the same month (Malhotra and Birks 2007, pp. 416-417). If N / n is not a whole number, rounding the interval can change the sample size: from 300 people with a rounded interval of 4 instead of 3.75, you get only 75 people instead of 80 (300 / 4 = 75). In that case, keep the fraction: choose a uniformly distributed random start u between 0 and 3.75 (excluding 3.75 itself), then the positions u, u + 3.75, u + 7.5 and so on, each rounded down, plus 1. This gives you exactly 80 people, and each has the same chance. Or use simple random sampling.

Stratified random sampling

You divide the population into groups that don't overlap, for example departments, locations or customer segments. Then you draw a separate random sample in each stratum (Jacob et al. 2019, p. 77). The strata should be as similar as possible internally and as different as possible from each other. The stratification variable should be related to what you're measuring. Malhotra and Birks (2007, pp. 417-418) advise rarely using more than two variables and no more than six strata.

With proportional stratification, each subsample matches its stratum's share of the population. If production makes up 60% of the workforce, it also makes up 60% of the sample. You can pool the responses without weighting (Schumann 2019, pp. 94-95).

With disproportionate stratification, the shares deliberately differ, usually to give small groups more cases. Jacob et al. (2019, p. 77) write that disproportionate stratification dominates in practice, because it makes small groups analyzable without enlarging the whole sample. If your main goal is to compare groups, you'll often draw the same number per stratum (Malhotra and Birks 2007, p. 418). The price: for statements about everyone together, you have to weight, otherwise the small groups count too much (Schumann 2019, p. 94). The example in the next section shows how.

Don't confuse stratification with quotas. Both specify characteristics, but stratified sampling uses random selection within each group. With quota sampling, selection depends on who is available (Jacob et al. 2019, p. 78, n. 11).

Cluster sampling

In cluster sampling, you don't draw individuals but whole groups, for example school classes, branches or courses, and then survey everyone in them (Schumann 2019, p. 92). This pays off when there's no list of individuals but there is a list of groups, or when the individuals are widely scattered and you want to save travel costs (Jacob et al. 2019, pp. 78-79).

The catch is the cluster effect. Groups are often internally similar. If you draw five courses at a university, you may end up, by chance, with only psychology and business (Jacob et al. 2019, p. 79). The more and the smaller the clusters you draw, the lower the risk (Schumann 2019, p. 93). The analysis has to account for this effect, or the results look more precise than they are. Malhotra and Birks (2007, p. 422) therefore call cluster samples imprecise and difficult to analyze.

Strata or clusters? Stratification is meant to increase precision, clustering is meant to reduce costs (Malhotra and Birks 2007, p. 419). That's why strata should be internally similar, and you draw from every stratum. Clusters should be as mixed as possible internally, and you draw only some of them (Jacob et al. 2019, pp. 77-79).

Multistage sampling

Multistage sampling involves random selection at several successive stages. The result of each stage is the list for the next one. The best-known example from Germany is the ADM design, named after the German association of market and social research agencies (Arbeitskreis Deutscher Markt- und Sozialforschungsinstitute): first areas, then households, then one person in each household. According to Jacob et al. (2019, p. 80), it has been used for nearly all representative population surveys in Germany since 1978, which makes it the German standard. You'll rarely need anything like it for a university or company survey. But it shows how much effort lies behind the word "representative."

My own compilation based on Jacob et al. 2019, pp. 74-80, Malhotra and Birks 2007, pp. 415-422, and Schumann 2019, pp. 92-95. You'll find a similar overview in my course "Principles of Survey Research" (slide 134).

Worked example: a stratified sample in three departments

An invented but typical example. A company with 2,000 employees runs a large employee survey with everyone once a year. In between, management wants a short pulse survey without burdening everyone every quarter. The plan is 240 responses. How to arrive at a number like that is explained in the article Calculating sample size. Here we take it as given and pretend that everyone drawn has responded. One of the questions is:

"Overall, how satisfied are you with your work?"
very dissatisfied (1) · somewhat dissatisfied (2) · neither satisfied nor dissatisfied (3) · somewhat satisfied (4) · very satisfied (5)

The workforce is spread across three departments: 1,200 in production, 600 in administration, 200 in sales.

Option 1: proportional. Each department gets its share: 144 from production, 72 from administration, 24 from sales. You can pool the responses without weighting. For sales on its own, though, 24 responses is tight. Jacob et al. (2019, p. 72) give a rule of thumb of at least 20 to 25 cases per subgroup if you still want to say something about that group. If a few drop out, you're below that.

Option 2: the same number per department. Management wants to compare the departments and therefore draws 80 people at random from each one. This is a stratified random sample with disproportionate allocation. In sales, two out of five people are drawn, in production only one in fifteen.

Invented example with 80 respondents each, response options 1 to 5. Production: 8, 16, 28, 20, 8. Administration: 2, 8, 22, 32, 16. Sales: 1, 4, 15, 36, 24.

What happened? The satisfied sales team makes up a third of the sample, but only a tenth of the workforce. If you pool the 240 responses without weighting, the company in this example appears more satisfied: 56.7% instead of the weighted estimate of 46.5% satisfied. That's not a random error, it's a calculation error: the unweighted value ignores the different selection probabilities. Which direction it goes in depends on which group is overrepresented and how it responds.

The solution is design weighting. Each response gets a weight proportional to the inverse of its selection probability (Schumann 2019, p. 95). A person from production stands for 15 employees, one from sales for 2.5. So that the weighted number of cases adds up to 240 again, you divide all weights by N / n, here 2,000 / 240 = 8.33. This gives the weights 1.8, 0.9 and 0.3. The weighted mean of 3.32 is the same value you get if you weight the three department means by 60%, 30% and 10%. You can use this to check your calculation.

Equal-sized groups make department comparisons easier, but may reduce the precision of the overall estimate. Using a common approximation developed by the statistician Leslie Kish, the 240 responses here are roughly equivalent to responses from 174 people selected using proportional allocation, assuming responses vary similarly within departments and ignoring the different sampling fractions. The sampling fraction is particularly large in sales, where two out of five employees are selected. This approximation is therefore a guide, not a measure of the actual precision. When responses vary similarly within departments, proportional allocation is usually better for estimating the overall result, while equal allocation is better for comparing departments.

Two rules from this example apply in general:

• Report department-level results without weighting in this example. Within each department, every person had the same selection probability, so the sales mean is the sales mean. You only need weights for statements about everyone together.
• When in doubt, report separately rather than weight incorrectly. Jacob et al. (2019, p. 78) warn that wrong weights can distort the results even more. If you're unsure, report the groups individually.

How to describe it in your report: "The survey used a stratified random sample with disproportionate allocation from the staff list (as of March 1): 80 employees each from production, administration and sales, drawn using random numbers. Department-level results are unweighted. Overall results use design weights based on selection probabilities."

Design weighting only corrects what you caused yourself through the draw. If only the satisfied people in production respond, no design weight will help. That's a nonresponse problem. Schnell (2019, p. 171) notes that weighting doesn't solve it, it only conceals it. You can read more about weighting after the survey in the article Representative sample.

Non-random methods: what they can and can't do

Non-random methods (also called non-probability sampling) aren't automatically unscientific. Jacob et al. (2019, p. 74) consider them scientifically justifiable if they're theoretically grounded and carried out transparently. Sometimes there's no alternative at all, because no list exists or the group can't be captured in any register (Jacob et al. 2019, p. 86). What matters is that you know which method you're using and call it by its name.

Quota sampling

You specify how many people with certain characteristics should be included, for example by age and gender as in official statistics. Who gets surveyed within these targets is decided by the interviewers' convenience or judgment (Malhotra and Birks 2007, pp. 412-413). In an open online survey, the order decides: whoever fits and responds first counts, until the quota is full. Among non-random methods, quota sampling is the standard, especially in commercial market and opinion research (Jacob et al. 2019, p. 88). It's quick and cheap. But the sampling error can't be determined (Malhotra and Birks 2007, p. 414). Generalizing to the population requires the usually untested assumption that the quota sample behaves like a random sample (Schumann 2019, p. 98). Why matching quotas don't guarantee representativeness, and what role the combination of quotas plays, is explained in the article Representative sample.

Purposive sampling

You select people according to well-founded criteria, for example typical cases or experts (Jacob et al. 2019, p. 87). Malhotra and Birks (2007, p. 412) call this "judgemental sampling" and observe it frequently in B2B market research, because the target group there is often small. If you survey the twelve purchasing managers of your most important customers, that's purposive sampling. Your results apply to these twelve. Any generalization is only as well supported as your reasons for selecting those participants.

Convenience sampling

You survey whoever happens to be there: fellow students, the followers of the company's social media profile, people who click a link on your website. Malhotra and Birks (2007, p. 411) explicitly name some email and internet surveys as examples of these "convenience samples." They're the cheapest and quickest solution. At the same time, participants select themselves, and the sample doesn't stand for any definable population.

How strictly this is judged varies. Jacob et al. (2019, p. 86) call such samples arbitrary or haphazard selections and consider them completely unsuitable for scientific surveys. Malhotra and Birks (2007, p. 412) are milder: they're fine for pretests, pilot studies and initial exploration, but not for descriptive statements about a population. What you can always do: define whom you actually mean, and document where and when you invited people.

How to describe a convenience sample: "Participants were recruited through a link on the website and on the company's social media profile (fieldwork March 3 to 17). The results describe the respondents and can't be generalized to all customers." You'll find wording for a thesis in the article Representative sample.

Snowball sampling

You start with a few people who belong to the target group and ask them to name others or pass the questionnaire on. This creates waves of referrals (Malhotra and Birks 2007, p. 414). The method suits rare or hard-to-reach groups that you have no direct access to (Jacob et al. 2019, pp. 87-88). Even if the first people were chosen at random, the result isn't a random sample, because the people who get referred are usually similar to those who refer them. By the way, if you ask friends or colleagues to forward the link, you're using exactly this method. How to recruit participants this way and others is explained in Where do I find respondents for my survey.

Online access panels and river sampling

An access panel is a pool of people who have agreed to take part in surveys regularly. Jacob et al. (2019, p. 92) stress that a panel isn't a sampling method in its own right: it can be based on random selection, quotas or self-registration. With panels recruited online, every stage acts as a filter, from internet access to registration to responding (Schnell 2019, pp. 285-286). What this means for representativeness is covered in the article Representative sample, and how to work with a panel provider in Where do I find respondents for my survey.

In so-called river sampling, participants are directed straight into a survey through ads, pop-ups or referrals. Schnell (2019, p. 286, n. 12) remarks that the name suggests a defined or even calculable method, which isn't the case. Treat such samples like convenience samples with a particularly large number of self-selection stages.

My own wording. Based on Jacob et al. 2019, pp. 73-74 and 92-93, Malhotra and Birks 2007, pp. 411-414, Schumann 2019, pp. 97-98.

By the way, a convenience sample can also answer a question clearly. Schumann (2019, p. 97) points out that a claim meant to hold without exception can be refuted with just a few counterexamples, and that doesn't require random selection. If someone claims "nobody reads privacy notices," a handful of people who do is enough to refute it.

Which sampling method fits your survey?

Malhotra and Birks (2007, p. 421) base the choice on the type of research. If you want to explore, a non-random selection often gets you further. If you want to estimate shares or means for a whole group, you need a random sample. For the four most common situations, this means:

Thesis. You usually have no list and little time. Then it'll be a convenience or snowball sample, and that's fine as long as your research question fits it. If you're exploring a field or checking whether a relationship occurs at all in the group surveyed, it's justifiable as long as you describe the method openly and phrase results as tendencies in this group. It isn't sufficient for estimating population proportions or claiming that a relationship holds in the population (Malhotra and Birks 2007, p. 412). If your university gives you access to a directory of all students in a program, you can draw at random from it and have a genuine random sample. How to present this in your methods section is explained in How to write my thesis.

Employee survey. In small and medium-sized companies, you usually invite everyone. A sample may include too few people from small teams, yet team-level results are precisely what makes an employee survey useful. For a short pulse survey between two large rounds, a sample can make sense. Linke (2018, pp. 33-34) suggests random or quota selection for this and stresses that it's the separate analysis by area that determines the required size. If you want to compare areas, stratify by area, as in the example above.

Customer survey. Your customer database is a sampling frame that researchers studying the general population can only dream of. Jacob et al. (2019, p. 126) consider online surveys well suited precisely when they're based on such an address list. Use simple random sampling or stratify by customer segment, for example individual and business customers. Check coverage first: who has no email address on file, who has objected to being contacted? Otherwise these groups will be missing without anyone noticing.

General population. Statements about the population of a whole country require a random sample of that population, usually multistage or drawn from population registers. That's laborious and expensive (Jacob et al. 2019, pp. 80-85). A quota panel built on self-registration is no substitute (Jacob et al. 2019, p. 93). If you need a statement like that, commission a research institute and ask for a written description of its sampling method.

The decision tree is deliberately simple: it doesn't replace expert advice for a large study, but it gives your method a name before anyone asks.

How to put your sampling method into practice in QUESTIONSTAR

You make the selection beforehand in your list, for example in Excel or in your statistics software. If you work with an AI assistant connected to QUESTIONSTAR via the MCP interface, it can draw the random sample from your list, also separately for each stratum, and add the selected people as survey respondents, including an extra field such as department. It only sends invitations if you explicitly ask it to. After that, QUESTIONSTAR supports you at the points where samples most often fail in practice.

• Random sample from a list. You import the people drawn as a contact list from a file (from Excel, for example). You can include extra fields such as department. Each person receives a personal link by email, and reminders go only to those who haven't responded yet. This also works if you set up the survey as anonymous.
• Stratified sample. Draw separately for each stratum. Include the stratum as a question in the questionnaire, for example "Department," or have it prefilled from the contact list field. That way the stratum is recorded with every response: you can filter by it in the reports, and it's there in the export for weighting. In very small departments, such a question can make individuals identifiable, so combine those departments. For weighted overall values, export the raw data as an SPSS, CSV or XLSX file and apply the design weights there.
• Quotas. You set a maximum per answer option or respondent profile and decide what happens when a quota is full: go to the end of the survey, take a different path, or show a friendly rejection message. You can see the status of your quotas in the completions widget. If you work with a panel provider, QUESTIONSTAR redirects to the provider's quota-full URL once a quota is full. Keep in mind: a quota limits groups that are already responding. If a group is missing, you need to recruit people from that group directly.

Checklist: describing your sample in the methods section or report

• Population precisely named (who, where, which period)
• Sampling frame named and gaps described (who is missing from the list?)
• Method called by its name, for example "stratified random sample with disproportionate allocation" or "convenience sample via a link on the website"
• For random samples: how they were drawn (program, seed), how many per stratum
• For quotas: which characteristics, where the target values come from
• Invitees, respondents and response rate reported
• Weighting described, or the reason for not weighting given
• Statements in the results section or report match the method (see the table above)

Frequently asked questions about sampling methods

What is the difference between random sampling and quota sampling?

In a random sample, a random procedure decides who is surveyed, and every person has a known chance greater than zero. In a quota sample, only the shares of certain characteristics are fixed, and who is surveyed within the quotas depends on who can be reached and on the recruiter's judgment. That's why a sampling error can only be calculated for a random sample (Malhotra and Birks 2007, p. 414).

Is a stratified sample the same as a quota sample?

No. Both specify characteristics, but in a stratified sample, people are drawn at random within each stratum. It's a random sample, while a quota sample is a form of purposive selection (Jacob et al. 2019, p. 78, n. 11).

When do I need weighting?

For overall means or shares across everyone, you usually need design weights if the selection probabilities are unequal, for example because your strata deliberately don't match their share of the population. The weight is proportional to the inverse of the selection probability (Schumann 2019, p. 95). For results per stratum, you don't need weights, as long as everyone within the stratum had the same chance.

What is cluster sampling, in simple terms?

You draw whole groups, for example school classes, and survey everyone in them. This saves effort when there's no list of individuals. Because groups are often internally similar, the result is usually less precise than with a simple random sample of the same size (Jacob et al. 2019, p. 79, and Malhotra and Birks 2007, p. 422).

Is an online survey with a shared link a random sample?

No. Whoever sees the link and clicks it selects themselves. That's a convenience sample (Malhotra and Birks 2007, p. 411). It only becomes a random sample if you draw from a list and invite the people drawn personally, and even then only insofar as nonresponse doesn't distort the picture (Schnell 2019, p. 285).

Can I run significance tests on a convenience sample?

You can run them, but the precondition of random selection isn't met. Without additional assumptions, such samples don't allow conclusions about the population (Schumann 2019, p. 97). So report test results only under the explicitly stated assumption that the sample behaves like a random selection, and not as evidence about the population.

Sources

Jacob, R., Heinz, A. and Décieux, J. P. (2019). Umfrage. Einführung in die Methoden der Umfrageforschung [Surveys: an introduction to the methods of survey research] (4th ed.). Berlin/Boston: De Gruyter Oldenbourg. In German. pp. 72-85 (sample size, random selection), pp. 86-93 (non-random selection, access panels), p. 126 (online surveys).

Linke, R. (2018). Mitarbeiterbefragungen optimieren. Von der Befragung zum wirksamen Management-Instrument [Optimizing employee surveys: from survey to effective management tool]. Wiesbaden: Springer Gabler. In German. DOI 10.1007/978-3-658-17722-5. pp. 33-34 (partial and pulse surveys).

Malhotra, N. K. and Birks, D. F. (2007). Marketing Research. An Applied Approach (3rd European edition). Harlow: Pearson Education. pp. 407-422 (Sampling: design and procedures).

Marx, P. (2026). Principles of Survey Research. Online course, chapter 4, slides 122 and 134. QUESTIONSTAR. Slide 122 and slide 134.

Schnell, R. (2019). Survey-Interviews. Methoden standardisierter Befragungen [Survey interviews: methods of standardized surveys] (2nd ed.). Wiesbaden: Springer VS. In German. DOI 10.1007/978-3-531-19901-6. p. 171 (dealing with nonresponse), pp. 285-286 (samples in web surveys, river sampling).

Schumann, S. (2019). Repräsentative Umfrage. Praxisorientierte Einführung in empirische Methoden und statistische Analyseverfahren [Representative surveys: a practical introduction to empirical methods and statistical analysis] (7th ed.). Berlin/Boston: De Gruyter Oldenbourg. In German. pp. 83-85 (overview of sampling methods, population, frame population, inference population), pp. 92-98 (clusters, stratification, design weighting, purposive selection).

I hope this helps you describe your sampling method confidently, whether in a thesis or a report to management, and keep your conclusions within the limits of that method.

Related articles:

• Representative sample: what it is and how to check yours
• Calculating sample size: technical details explained simply
• Where do I find respondents for my survey

Dr. Paul Marx

Author: Dr. Paul Marx
Dr. Paul Marx is the founder and Managing Director of QUESTIONSTAR and was a professor of marketing at the University of Siegen.