What is sampling in research?

Sampling definition (and the 5 terms you must get right)

Sampling is the process of selecting a smaller group (a sample) from a larger group (a population) so you can learn about the population without measuring everyone. A full-population study is often impractical, so researchers sample to make the work feasible while still producing useful conclusions.

The five terms below do most of the work in any methodology chapter. Get them wrong and every later decision inherits the mistake.

• warning
  Population: the full group you want to describe, written with boundaries and a time window. "All final-year nursing students at the three campuses, enrolled as of 1 March" is a population. "Nurses" is not.
• warning
  Target population vs accessible population: who you want to describe versus who your recruitment and tools can actually reach. Naming the gap between the two is a strength in a write-up, not a weakness.
• warning
  Sampling frame: the list or mechanism you draw from. A student register, a patient register, a customer database, a village household listing, a voter roll. Frame problems create coverage error, and no method repairs them.
• warning
  Sampling unit: what actually gets selected. A person, a household, a clinic, a class, a shift, an order. In cluster and multistage designs the unit at the first stage is a group, and the group is called the primary sampling unit.
• warning
  Sampling design: the whole plan written down. Frame, method, how many units, how they are contacted, and what you will do about the people who do not answer. "Sampling design" is what an examiner or a reviewer is actually asking for when they ask about your sampling.

For a short official definition of the same idea, see the U.S. Office of Research Integrity training note on sampling.¹ If you are placing sampling inside a wider study plan, our guide to survey research design covers the steps either side of it.

Why sample at all, and when a census is the better answer

The need for sampling is practical before it is statistical. Four reasons come up again and again.

1. Cost

   Every extra respondent costs money, staff time, or goodwill. A well-drawn sample of 600 can answer the same question as a badly run attempt at 60,000.

2. Speed

   Decisions have deadlines. A sample gives you an answer inside the window where the answer is still useful.

3. Feasibility

   Some populations cannot be enumerated at all: passers-by at a market, people who abandoned a service, informal traders. Others are effectively infinite, or measuring a unit destroys it.

4. Quality

   This one surprises people. Concentrating effort on a smaller group lets you chase non-responders, train interviewers properly, and check the data. Effort per unit is usually worth more than units.

A census measures everyone instead. It is the better choice when the population is small enough to reach, when contact details already exist, or when every member has a right to be asked. An employee survey in a 200-person organisation is a census question, not a sampling question. So is a class evaluation. When the population is small, sampling buys you almost nothing and costs you the ability to report by team.

What a good sample actually looks like

Textbooks describe a good sample as "representative", which is not an operational test. These four properties are.

• Coverage: everyone in the target population could in principle have been selected.
• Controlled selection: who ends up in the sample is decided by a rule, not by who volunteers or who is nearest the door.
• Enough units in the smallest group you must report: an overall number can look healthy while the group you actually care about rests on 11 responses.
• Describable: you can write down exactly how it was drawn, and someone else could repeat it.

For a plain-language account of how samples support population estimates, the U.S. Census Bureau overview of survey sampling and inference is a good starting point.²

Your sampling frame sets the ceiling

A sampling frame is the concrete list or mechanism you draw from. It is the most under-discussed part of sampling and the place where most real studies quietly go wrong, because the frame decides who was even eligible to be selected.

Frames are ordinary objects. A student register, a patient register, an HR roster, a customer database. A list of villages with household counts, a voter roll, a run of orders, a list of clinics in a district. Notice that some frames list people and some list places. That distinction decides which methods are open to you.

Audit the frame before you draw, not after. Count the rows. Compare that count with any independent total you can find, such as an HR headcount or an enrolment figure. Look for the segment you would be most embarrassed to have missed, and check that it is present. Deduplicate. Write down what the frame excludes, because that sentence belongs in your limitations either way.

The 9 sampling methods compared

Sampling methods split into two families. Probability sampling uses a random mechanism, so every unit in the frame has a known, non-zero chance of selection. Non-probability sampling does not, so selection probabilities are unknown. Everything else follows from that one distinction.

The second column is the one most guides leave out, so it comes first here. Only a probability design supports a margin of error, a confidence interval, or the sentence "plus or minus 5 points". If a method says no there, you can still learn a great deal from it. You just cannot publish a precision figure alongside the result. Precision itself belongs to a different conversation, covered on our sample size guide.

Reading the table: the first five methods are probability designs and the last four are not. That single split, not the individual method names, is what decides whether a precision figure is honest.

Probability sampling methods, one by one

Probability sampling means every unit in the frame has a known, non-zero chance of being selected, and a random mechanism does the selecting. That is the whole basis for saying anything numerical about the population. Five designs cover almost every real study. Clinical and health research reviews cover the same family under slightly different labels, and it is worth knowing why before you meet both: each of the two cited here folds multistage sampling into cluster sampling rather than listing it separately.³⁴

Simple random sampling

Simple random sampling gives every unit on the list exactly the same chance of selection. You number the list, generate random numbers, and take the units they land on.

Worked example. A university register holds 12,000 enrolled students and you need 400 responses. Number the register 1 to 12,000, generate 400 unique random numbers in that range with a spreadsheet or a statistics package, and invite those students. If you expect a 40 per cent response rate, draw and invite 1,000 so that 400 completed responses is realistic.

Where it fails. With 400 responses spread across a 12,000-student register, a department of 300 students contributes about 10 people. That is not enough to report on. If any subgroup has to appear in your results, simple random sampling is the wrong tool and stratified sampling is the fix.

Write down your randomisation. Software used, the seed if there was one, and the date. That single sentence is what makes the draw reproducible, and reviewers do ask for it.

Systematic sampling

Systematic sampling takes every kth unit from an ordered list after a random start. Divide the list size by the number you need to get the interval k, pick a random number between 1 and k as your starting point, then count forward.

Worked example. A district hospital records 9,000 outpatient visits in a quarter and wants 600 exit interviews. The interval is 9,000 divided by 600, so k equals 15. Pick a random start between 1 and 15, say 7, then interview visit 7, 22, 37, 52 and so on. A field team can follow that rule without a laptop, which is the real reason this method survives.

Where it fails. Periodicity. If the list has a repeating cycle that lines up with k, you sample the same slot every time. A visit list ordered by clinic day with a 7-day cycle and an interval of 7 would give you one weekday and nothing else. Check the ordering before you fix k, and shuffle the list if the order relates to what you are measuring.

Stratified sampling

Stratified sampling divides the population into non-overlapping groups called strata, then draws a random sample inside every stratum. It is the most useful upgrade available to most studies, because it guarantees that each group you must report on actually appears.

Worked example with the arithmetic shown. An organisation has 4,200 staff across six provinces and needs 600 responses.

• Proportional allocation. A province with 1,400 staff holds 1,400 divided by 4,200, which is one third of the workforce, so it gets one third of 600, which is 200 responses. A province with 210 staff gets 210 divided by 4,200 times 600, which is 30 responses. Every province is represented in proportion to its size, and no weighting is needed at analysis.
• Disproportionate allocation, also called oversampling. Thirty responses is too few to report on that small province separately. So you deliberately take 80 from it instead, and reduce the largest province to compensate. Now selection probabilities differ between provinces, so you must weight the data back to true proportions before quoting any organisation-wide figure. Skipping that step is the single most common stratified sampling error.

Choosing strata. Pick variables that relate to the outcome or to who is likely to respond, and that already exist on your frame. Region, site, role, tenure band, course, plan type and age band are the usual candidates. Strata that have nothing to do with the outcome add administration and buy no precision. If the grouping variable is missing from your frame, you may need to collect it inside the questionnaire instead. Our guide to demographic survey questions covers how to ask for those variables without losing respondents.

Cluster sampling

Cluster sampling randomly selects whole naturally occurring groups and then surveys everyone inside the selected groups. You need a list of the groups, not a list of the people, which is why it exists.

Worked example. A district has 180 primary schools and no central list of pupils. You randomly select 24 schools and survey every Grade 8 pupil in each one. You have made 24 school visits instead of 180, and you never needed a pupil register to start.

What you pay for it. Pupils in the same school resemble each other more than they resemble pupils across the district. So each extra pupil in a school you have already picked adds less new information than an independent pupil would. Your effective number of responses is smaller than your actual number of responses. That inflation is the design effect, and it has to be built into the plan rather than discovered at analysis.

Where it fails. Too few clusters. Six schools chosen at random can differ wildly from each other, and the result swings on which six you happened to draw. More clusters with fewer people in each is almost always better than the reverse.

Multistage sampling

Multistage sampling selects in two or more stages: groups first, then units inside the selected groups, and sometimes a third stage inside those. Almost every national household survey in the world is built this way.

Worked example. Select 120 wards from the national list of wards. Inside each selected ward, a field team lists the households on the ground, then randomly selects 20 of them. Inside each selected household, one adult is chosen at random rather than left to whoever opens the door. That is three stages, and the last one matters: letting the household decide who answers reintroduces exactly the selection bias the design was built to avoid.

What you pay for it. Different people end up with different selection probabilities, so the analysis needs weights, and the variance calculation has to account for the clustering at every stage. This is a real design that needs someone who can carry out a weighted analysis. If nobody on the project can, choose a simpler design rather than an unweighted multistage one.

Stratified vs cluster sampling: the difference in one line

These two get confused constantly, because both start by dividing the population into groups. The difference is what you do next.

Non-probability sampling methods, one by one

Non-probability sampling means selection probabilities are unknown, because no random mechanism decided who was chosen. It is the right choice in four situations. You cannot build a usable frame. The group is hidden. Speed matters more than precision. Or the question is about meaning rather than size.

These methods are not second-class. They are answering a different question. The failure is never using them, it is using them and then writing a sentence that only a probability sample could support.

Convenience sampling

Convenience sampling, also called availability sampling, recruits whoever is easiest to reach. A pop-up on a website, a link in a newsletter, students in your own class, people passing a stall in a market.

Example. A shop places a feedback link on the receipt. Everyone who answers has bought something, was carrying a phone, had time, and had a reason to bother. Each of those filters removes a different part of the customer base.

Honest use. Pilot testing a questionnaire, generating hypotheses, finding out what language customers use, checking that a question is understood. All excellent uses. None of them require a population estimate.

Quota sampling

Quota sampling sets recruitment targets for named groups and fills them. Forty rural and sixty urban. Fifty women and fifty men. Twenty in each of five age bands.

Example. A market research team needs 500 responses matching the national split on age, gender and region, and has no list of the population. They recruit until each cell is full.

Why it is still not probability sampling. The quotas control the shape of the sample, not the selection of the people in it. Whoever recruits still chooses, and people who are easy to approach are systematically different from people who are not. A sample can match the population perfectly on age and gender and still be badly wrong about attitudes, because age and gender were never the variables doing the damage.

Purposive sampling

Purposive sampling, also called judgement sampling, deliberately selects cases that meet stated criteria. Common variants are typical-case, extreme-case, critical-case, maximum-variation, and expert selection.

Example. A study of how a new payment system is used selects eight people. Two who use it daily, two who gave up on it, two who never took it up, and two support staff who handle complaints. That spread is chosen on purpose, and the reasoning is the method.

Honest use. Most qualitative research. The standard here is not randomness. It is whether you can say why these cases and not others, and whether the range of views is wide enough to support your claims.

Snowball sampling

Snowball sampling recruits a small number of participants who then refer others who fit the criteria, and those refer more.

Example. A study of informal cross-border traders begins with four participants identified through a local association, and each one introduces two more. There is no register of informal traders, so no frame-based method is available.

What it does to your sample. Homophily. People refer people like themselves. The sample drifts toward one social network and away from isolated members of the population, who are often the ones the research most needs to hear. Starting from several unconnected seeds rather than one reduces the drift but does not remove it.

How to make a non-probability sample defensible

Most guides stop at "you cannot generalise" and leave the reader stranded, because most readers are running exactly this kind of study. Here is the part that usually goes missing.

• Describe the recruitment, not just the method name. Where the link was posted, over what dates, who saw it, whether an incentive was offered, and how many were approached versus how many answered. This is what lets a reader judge the sample instead of guessing.
• Compare your achieved sample against a known benchmark and publish the comparison. If your organisation is 60 per cent field staff and your respondents are 25 per cent field staff, say so in a table. A stated skew is a finding. A hidden one is a defect.
• Know what weighting can and cannot do. Adjusting your data so the demographics match the population can correct visible imbalance. It cannot correct for people who differ from respondents on something you did not measure, and it never converts a non-probability sample into one that supports a precision claim.
• Change the sentence, not the statistics. Instead of "38 per cent of customers are dissatisfied, plus or minus 4 points", write "38 per cent of the 620 customers who responded reported dissatisfaction. Respondents were recruited by an in-app prompt over two weeks and are not a random sample of customers." The second sentence is shorter, more useful, and cannot be attacked.

Qualitative work has its own vocabulary for this. A protocol usually asks for a rationale rather than a numeric target. It should cover the intended number and the range of views, a distinction the University of Bath guide states concisely.⁵

Match the method to the situation

Six situations, six methods. Decide before you open the answer. This is the fastest way to find out whether the table above actually landed.

How to choose a sampling method

Most studies do not fail because someone picked the wrong named technique. They fail because the technique did not match the frame, the decision, or what the field team could realistically do. Three questions settle it almost every time.

1. Start with the claim you need to make

   Write the sentence you want to publish, before you choose anything. "Sixty-two per cent of our customers would recommend us" and "customers who leave describe three recurring problems" are different sentences that need different designs. If nobody will treat the result as a population figure, you have far more freedom than you think.

2. Audit the frame before you commit

   Can you actually list the population, or a close proxy? If not, can you list the groups they belong to? Answering this one question eliminates most of the nine methods immediately, which is why it goes second and not last.

3. Name the subgroups you are obliged to report

   If a result has to appear for each region, site or role, that drives the design through strata or quotas. It also drives how many responses you need in each group. Deciding this after fieldwork is not recoverable.

4. Check the design against the field reality

   Time, budget, contactability, language access, literacy, connectivity and incentive policy decide what is executable. A design a field team cannot follow becomes convenience sampling with a formal name attached, which is worse than choosing convenience sampling honestly. Inclusive recruitment planning reduces systematic underrepresentation.⁶

Where samples actually break: coverage, selection and nonresponse

Selection is the part everyone worries about and the part that goes wrong least often. The damage usually happens before selection, in the frame, or after it, in who bothers to answer. Both produce clean-looking spreadsheets and wrong conclusions.

The most famous sampling failure had 2.4 million respondents

In 1936 the Literary Digest predicted Alf Landon would take roughly 55 per cent of the vote. Franklin Roosevelt won about 61 per cent.⁷ The magazine had mailed 10 million ballots and got about 2.4 million back, a 24 per cent response rate that no national poll today comes close to.⁸ The poll put Roosevelt twenty points below his actual share, on a sample of two and a half million people.

The number that should be on every sampling page. A later reanalysis computed the sampling margin of error for the Digest's state-level estimates: it ranged from 0.18 to 2.3 percentage points, with a median of 0.6.⁸ A precision figure of plus or minus half a point, sitting on top of an error of twenty points. Nothing demonstrates more cleanly that a margin of error measures only one of the several ways a survey can be wrong.

And the standard explanation is only half right. Everyone is taught that the Digest sampled car owners and telephone subscribers, who were rich and Republican. One empirical test of that claim used a Gallup survey from May 1937. People who owned both a car and a telephone still broke for Roosevelt, 55 to 45. Among everyone who said they received a Digest ballot, Roosevelt led 55 to 44. The frame leaned Republican, but not nearly enough to produce a Landon landslide.⁷

The missing half is nonresponse. Among people who said they returned their ballot, Landon led 51 to 48. Among those who did not return it, Roosevelt led 69 to 30. The rough decomposition puts about eleven points of the error on the frame and about seven on differential nonresponse.⁷ Weighting the returns on a question the Digest had already asked, how you voted in 1932, would have produced the correct winner using arithmetic available in 1936.⁸

A high response rate is not the target you think it is

Response rates for Pew Research Center's own telephone polls fell from 36 per cent in 1997 to 9 per cent by 2012, and to 6 per cent by 2018.⁹ Meanwhile India's National Family Health Survey, conducted face to face, selected 664,972 households and found 653,144 of them occupied. It interviewed 636,699 of those, a 97.5 per cent response rate. Read that denominator carefully, because it is the most common place a response rate goes wrong: against the number selected rather than the number occupied, the same fieldwork reads as 95.7 per cent. The survey drew its sample from 30,456 clusters across all 707 districts.¹⁰

Those two facts sit in the same decade. Mode and fieldwork effort drive response rates far more than any global claim about survey fatigue does. If a guide tells you the average response rate is some single number, it is not describing a real quantity.

Now the harder finding. A meta-analysis of 235 nonresponse-bias estimates drawn from 30 published studies found the correlation between response rate and the size of nonresponse bias was 0.33. Squared, that means the response rate explains about 11 per cent of the variation in bias.¹¹ A follow-up across 59 studies designed specifically to measure nonresponse bias reached the same conclusion.¹²

The response rate is worth reporting and worth improving. It is not a certificate. What causes bias is whether the tendency to respond is tied to the thing you measure. A survey about volunteering suffers from that far more than a survey about broadband speed. Chase the correlation, not the percentage.

More responses cannot repair a broken frame

During the 2021 vaccination rollout, two very large surveys tracked United States vaccine uptake. One collected about 250,000 responses a week and overstated first-dose uptake by 17 percentage points against the official benchmark. The other collected about 75,000 every two weeks and overstated it by 14. A third survey of roughly 1,000 respondents, run to standard probability practice, produced reliable estimates.¹³

The same pattern shows up in ordinary survey work. Comparing three probability panels against three online opt-in sources across 28 benchmarks and 29,937 interviews, average absolute error was 5.8 percentage points for the opt-in samples and 2.6 for the probability panels. On the 25 of those benchmarks that were available for subgroups, the gap for adults aged 18 to 29 was 11.2 against 3.6.¹⁴

And a memorable diagnostic: in one opt-in survey, 12 per cent of adults under 30 said they were licensed to operate a nuclear submarine.¹⁵ The true figure is close to zero. If a panel can produce that, it can produce anything, and a larger order from the same panel simply buys more of it.

Who your frame misses, outside the US and UK

Most sampling guides quietly assume a Western frame: an email list, a landline, a national address register. Much of the world does not sample that way, and the coverage gaps have a measured shape.

• An online survey in a low or middle income country is a sample of the connected part of it, and that part skews male. Women in low and middle income countries are 14 per cent less likely than men to use mobile internet, which works out at roughly 235 million fewer women than men online.¹⁶
• Phone samples over-represent household heads. In World Bank phone surveys across four African countries, respondents were 52 per cent male in Uganda and 73 per cent male in Nigeria, against roughly half the adult population. In Uganda 6 per cent of phone respondents were aged 15 to 24, against 38 per cent of the adult population.¹⁷
• What to do about it. Combine modes rather than defending one. List households on the ground where no register exists. Field in the languages your population actually reads, and pre-test the translation rather than back-translating it. Then state the residual gap in your limitations, because it will still be there.

What weighting can and cannot fix

Weighting adjusts your data so the sample composition matches known population figures. It is genuinely powerful and it is routinely oversold.

The case for it. In 2016 the final University of New Hampshire poll had Clinton ahead by 11 points in a race she won by 0.4. It had not adjusted for education. The review committee concluded that this single adjustment would have removed essentially all of the error.¹⁸ One variable, correlated with both response and outcome, doing all the work.

The case against relying on it. When more than 30,000 online opt-in interviews were scored against 24 benchmarks, average estimated bias was 8.4 percentage points unweighted. Across every weighting procedure tested, including sophisticated matching approaches, none brought average bias below 6 points.¹⁹ Quadrupling the sample size bought a further 0.2 points.

The rule that follows is simple. Weighting corrects imbalance on the variables you weight on, to the extent those variables relate to the outcome. It does nothing about the ways respondents differ from non-respondents on things you never measured, and it never converts a non-probability sample into one that supports a precision claim.

A precision figure is a floor, not a ceiling

A margin of error quantifies sampling variability, and only that. It assumes the sample was drawn at random from the target population and that the remaining bias has been removed. Coverage error, nonresponse bias, measurement error and fraudulent respondents are all outside it.

The evidence is blunt. In the final fortnight of 2020, 348 state-level presidential polls reported a margin of error, and the average reported figure was 3.9 points. Measured against that average, 55 per cent of them missed the final margin by more than one margin of error, and 21 per cent missed by more than twice it.²⁰ These were modern polls with modern methods. Treat the margin of error as a floor on your uncertainty, never a ceiling.

Grade A means an official statistic, a meta-analysis, or a replicated finding. Grade B means one strong study whose conclusion has not yet been independently repeated.

Design effect: why clustered samples need more people

Almost every guide recommends cluster sampling for its cost saving and stops there. The saving is real, and so is the bill. People inside the same school, clinic, ward or store resemble each other. They resemble each other more than they resemble people picked at random from across the population. So each extra response from a group you have already picked carries less new information.

The design effect is the number that measures this. In words: how much wider your uncertainty gets compared with drawing the same number of people independently. For a design with equal-sized groups it is

design effect = 1 + (average group size - 1) x intraclass correlation

The intraclass correlation is how alike people inside a group are on the thing you are measuring. Zero means being in the same school tells you nothing about a pupil's answer. One means everyone in a school answers identically, and the school might as well be one person. Real values are small, but small values do a lot of damage once group sizes get big. Pooled across 48 national household surveys, the average intraclass correlation is about 0.06. It varies by what you measure. Fertility items run about 0.01 to 0.03, current contraceptive use 0.03 to 0.05, and medically delivered births as high as 0.22 in the national figures.²¹

If you have no estimate of your own, borrow a default carefully. The DHS sampling manual falls back on 1.5 for deft when a survey has no figure of its own.²² Deft is not the same quantity as the design effect in the formula above: it is the ratio of the two standard errors, where the design effect is the ratio of the two variances. You square it, so a deft of 1.5 is a design effect of 2.25. The manual calls the value a fallback rather than a typical one, and its own worked examples use 1.40 for modern contraceptive use and 1.22 for child vaccination.

Weighting adds its own component. A national survey that oversamples one region and then weights it back carries two design effects at once. A weighting effect of 1.22 and a clustering effect of 1.80 give an overall design effect of 2.20. So 5,000 interviewed households behave like 2,277 independent ones.²³

The design lesson. More groups with fewer people in each beats fewer groups with more people in each, almost every time. Forty schools of 24 pupils and 24 schools of 40 pupils cost roughly the same 960 responses, but the first has a smaller design effect and a defensible result. If you are planning how many responses you need in total, work that out on our sample size guide and then multiply by the design effect. Doing it the other way round is how clustered studies end up underpowered.

Three worked examples, start to finish

Method names are easy. The decisions around them are where studies actually go wrong. These three carry a real design from population statement through to the sentence you can publish.

Example 1: patient experience across a district

• Population. Adults who attended an outpatient consultation at any of the 46 public clinics in the district between 1 April and 30 June.
• Frame. No district-wide patient register exists. A list of the 46 clinics does, with monthly attendance counts.
• Method. Multistage. Select 18 clinics with probability proportional to attendance, then systematically sample every 8th exiting patient during two randomly chosen clinic days.
• Why not something simpler. A full patient list would allow stratified sampling, but it does not exist and cannot be built inside the timeline. A convenience sample at the three largest clinics would be quicker and would describe the three largest clinics.
• What must be reported. Selection probabilities differ by clinic size, so the analysis needs weights, and the interval needs widening for the design effect.
• The publishable sentence. "Among adults attending public outpatient clinics in the district during the second quarter, an estimated 71 per cent rated their wait acceptable, based on 1,240 exit interviews across 18 clinics selected with probability proportional to attendance."

Example 2: a dissertation study on a student population

• Population. Students enrolled in the Faculty of Commerce in the current academic year: 3,100 across three years of study.
• Frame. The faculty enrolment register, which carries year of study and programme.
• Method. Stratified random sampling by year of study, proportional allocation, target 350 completed responses.
• The arithmetic. First year holds 1,300 students, which is 42 per cent of 3,100, so it gets 42 per cent of 350, which is 147 responses. Repeat for each year. Because allocation is proportional, no weighting is required.
• Response planning. At an expected 35 per cent response rate you need to invite roughly 1,000 students to land 350 completed responses, and you should invite them stratum by stratum so a weak response in one year is visible while there is still time to act.
• Limits to state. Students who have withdrawn but remain on the register are overcoverage. Students enrolled after the register was pulled are undercoverage. Both belong in the limitations paragraph.

Example 3: customer feedback for an online store

• Population. Customers who completed an order in the last 90 days.
• Frame. The order table. It is complete for online orders and complete for guest checkouts, but a customer with two accounts appears twice.
• Method. Deduplicate to one row per customer, then systematic sampling of every 12th customer after a random start, running continuously rather than as one batch.
• The trap this avoids. Inviting only newsletter subscribers is faster and is a convenience sample of your most engaged customers. It will make satisfaction look better than it is, reliably, every quarter.
• A note on employees. The same organisation surveying its own 180 staff should not sample at all. That population is small, reachable, and everyone has a right to be asked. Run a census, and if participation is thin, treat that as a finding rather than a sampling problem.

How to write the sampling part of your methodology

This is the step most guides skip, and it is the one a supervisor, reviewer or client actually reads. A sampling write-up has to answer five questions in order: who, from what list, chosen how, how many, and what you are therefore allowed to claim.

Fill in the blanks below. It is deliberately plain: examiners are not looking for elegant prose here, they are looking for a design they can evaluate.

The same study, written weakly and written well

Weak. "A random sample of students was taken and 350 responses were collected."

Nothing here can be checked. "Random" is doing work it has not earned, there is no frame, no denominator, and no statement of who was excluded.

Strong. "The target population was the 3,100 students enrolled in the Faculty of Commerce in 2026. The sampling frame was the faculty enrolment register extracted on 3 March 2026. Stratified random sampling by year of study with proportional allocation was used so that each year could be reported separately. One thousand students were invited and 350 completed the questionnaire, a response rate of 35 per cent. No weighting was applied because allocation was proportional. Students who withdrew after the extract date remained on the frame and could not be identified, which is a minor source of overcoverage."

The second version is longer, and every extra sentence is a sentence a reviewer no longer has to ask about.

Sampling plan checklist

A written sampling plan prevents most avoidable errors, because it forces the decisions to happen before fieldwork rather than during analysis. Eight points, one page.

Once the plan is settled, the instrument decides whether the sample was worth drawing. A perfect sample cannot rescue a leading or double-barrelled question, so it is worth reading through survey question examples before you field anything.

Frequently asked questions

Related reading

Sampling is one decision inside a larger study. These guides cover the decisions either side of it.

• Survey research design puts sampling in sequence with the rest of a study, from defining the population to launching.
• How to determine sample size covers the precision arithmetic this page deliberately hands off, including the formulas.
• How to write survey questions covers the measurement side, which no sampling design can rescue.
• Questionnaire design principles covers how a whole instrument is assembled once the questions exist.
• Response bias goes deeper on the difference between the people you selected and the people who answered.
• Demographic survey questions is useful when your frame lacks the variable you need to stratify on.
• Statistical significance covers what to do with the numbers once the sample is in.
• Quantitative data analysis covers the analysis steps that follow, including where weights belong.

How this guide was written, and what it does not cover

Who wrote it. Michael Hodge, who works on survey methodology, question design and data quality at SuperSurvey, and writes the practical guides in our Survey Learning Center.

What it is based on. Definitions follow the methodological review and the university research guidance cited below, which is where the technical terms are actually defined. The official research-integrity and statistical-office pages are listed as short plain-language starting points, not as the source of those definitions. Method descriptions follow the same reviews and guidance. Every number in the evidence section comes from a named primary source listed below. Each claim also carries a confidence grade, so you can see which findings are settled and which rest on a single study. The design effect figures in the chart and table are our own arithmetic. We used the standard formula, with the assumptions printed in the caption, so you can redo them or argue with them using your own cluster sizes.

What it does not cover. This guide is about sampling for surveys and observational studies. It does not cover experimental design or random assignment beyond telling them apart. It also leaves out sampling for acceptance testing, auditing and signal processing, which use the same word for different things. Respondent-driven sampling, time-location sampling and adaptive cluster sampling are real methods that specialists use for hidden populations, and they are out of scope here. The intraclass correlation figures come from household surveys in low and middle income countries. Treat them as illustrative, not as values to import into a different setting.

Where the numbers are weakest. The Literary Digest reanalysis rests on one 1937 survey and one later reanalysis, and both authors flag their own uncertainty; it is graded B for that reason. Polling error figures come from United States election polls, which are unusually well studied and not necessarily representative of survey error in general.

Disclosure. SuperSurvey sells survey software. Nothing on this page is for sale, there is no sign-up gate on any part of it, and no organisation sponsored or reviewed this guide. Where we link to our own pages it is because they cover the next decision, not because a link earns us anything.

References