Guides  /  sampling

Sampling methods

How you select people decides what your results mean, and no amount of analysis afterwards can repair a bad selection. This guide covers the probability methods and the non-probability ones, works through stratified allocation with real arithmetic, shows what cluster sampling costs you in precision, and is honest about when stratifying buys almost nothing.

Elaine Halliburton Written and reviewed by Elaine Halliburton, Professor of Applied Statistics
Updated 4 November 202614 min read
What are the main sampling methods?

They divide into two families. Probability methods give every member of the population a known, non-zero chance of selection: simple random, systematic, stratified, cluster and multistage. Because the selection probability is known, you can calculate how far your estimate is likely to be from the truth. Non-probability methods, including convenience, purposive, quota and snowball sampling, do not. They are often the only practical option and are entirely legitimate, but they cannot support a confidence interval that means what it appears to mean.

Before any method

The sampling frame comes first

The frame is the list you sample from. A patient register, a staff database, an electoral roll, a customer file. Every property of your sample is inherited from it, which is why the frame deserves more attention than the method.

✓Undercoverage. People in the population who are not on the list. A landline frame misses mobile-only households, who are younger. Nothing in the analysis can detect this.
✓Overcoverage. Entries that should not be there, such as people who have left or died.
✓Duplication. Anyone listed twice has double the chance of selection, which quietly weights them.
✓Currency. A register updated annually and sampled in month eleven is a picture of last year.
The target population and the frame population are different things

Your conclusions apply to the frame, not to the population you had in mind. State both and describe the gap. "All adults in the borough" and "all adults registered with a GP in the borough" are not the same group, and the people in the difference are usually the ones the study most wanted to reach.

Known selection probability

The probability methods

MethodHow it worksUse whenCost
Simple randomEvery member equally likely, drawn at randomYou have a complete frame and the population is reachableCan be expensive and geographically scattered
SystematicEvery kth member after a random startThe frame is ordered and longDangerous if the list has a cycle matching k
StratifiedSplit into groups, sample within eachGroups differ on the outcome and you know who is in whichNeeds the grouping variable on the frame
ClusterSample whole groups, measure everyone insideThe population is naturally grouped and travel is costlySubstantial loss of precision
MultistageClusters, then units within themNational surveysComplex weighting and variance estimation

Systematic sampling, and its trap

The interval is the population divided by the sample.

k = N / n = 10,000 / 600 = 16.67, so k = 17

Pick a random start between 1 and 17, then take every 17th record. It is quick, it needs no random number generator at the point of selection, and in practice it is often easier to audit than a random draw because anyone can reproduce it from the start value.

It fails in one specific way. Taking every 10th shift from a rota where shifts cycle weekly will select the same weekday every time. The method is sound; the problem is an interval that resonates with a pattern already in the list. Check how the frame is ordered before choosing k. If the order is arbitrary, systematic sampling is a perfectly good substitute for simple random. If the order relates to the outcome, say a patient list sorted by admission date, it behaves like an implicit stratification and usually improves precision slightly.

The finite population correction

Standard error formulas assume you are drawing from an effectively infinite population. When your sample is a large fraction of a small population, that assumption is pessimistic and you are overstating your uncertainty.

fpc = √((N − n) / (N − 1))

Multiply the standard error by this factor.

PopulationSampleFractionCorrection factor
10,0006006%0.970, a 3% reduction
40020050%0.708, a 29% reduction

At 6% it is barely worth applying. At 50% it is substantial and ignoring it means reporting intervals almost a third wider than they should be. The usual rule is to apply it once the sample exceeds about 5% of the population, which happens more often than people expect in audits of single departments and surveys of small professional bodies.

Unknown selection probability

The non-probability methods

MethodHow it worksLegitimate use
ConvenienceWhoever is availablePiloting an instrument, early exploratory work
PurposiveDeliberately chosen for a characteristicQualitative work where you want information-rich cases, not representativeness
QuotaFill preset counts per subgroup, selection within them non-randomMarket research against time and budget limits
SnowballParticipants recruit othersHidden or hard-to-reach populations with no frame

None of these is a lesser method. Purposive sampling is the correct choice for most qualitative research, where maximum variation or extreme cases tell you more than a random draw would. Snowball sampling may be the only way to reach a population that no list describes.

What you give up is one specific thing

A confidence interval assumes a known probability of selection. Without it the interval still computes and still prints, but it no longer describes how far your estimate might be from the population value. Report descriptive statistics and be explicit that inference to a wider population is not supported. A quota sample of 1,000 is not a random sample of 1,000, however similar the demographic profile looks.

Quota sampling is the one most often mistaken for probability sampling, because matching the population on age and sex makes it feel representative. It controls the variables you quota on and nothing else, which is the same limitation that applies to weighting after the fact.

The four purposive variants worth naming

VariantSelect forAnswers
Maximum variationThe widest spread of the characteristicWhat is common across very different cases
HomogeneousA narrowly defined shared positionDepth within one group, typical of focus groups
Extreme or deviant caseThe outliersWhat happens at the edges, and why
Critical caseThe case where the mechanism is clearestIf it does not hold here, it holds nowhere

Naming the variant in a methods chapter is worth doing. "Purposive sampling" alone reads as a label; "maximum variation purposive sampling, to capture the range of departmental practice" reads as a decision. Examiners notice the difference.

How many for a qualitative study

There is no calculation, and anyone who offers you one is overselling. The working answer is information power: enough cases that further interviews stop producing new codes. Commonly cited ranges are 6 to 12 for a narrowly defined homogeneous group and 20 to 30 for a broader interview study, but these are conventions rather than thresholds.

Say how you judged it. "Recruitment continued until two consecutive interviews produced no new codes, which occurred at 14" is a defensible sentence. "Saturation was reached" on its own is not, because it describes a conclusion without describing how you arrived at it.

Worked example one

Stratified allocation

A survey of 10,000 businesses, sample of 600. Three size strata, and the outcome varies by size.

StratumNExpected pProportionalNeyman
Small6,0000.22360334
Medium3,0000.45180200
Large1,0000.606066

Step 1. Proportional allocation

Each stratum gets a share matching its share of the population.

Small: 600 × 6,000 / 10,000 = 360

Step 2. Neyman allocation

Sample more heavily where there is more variability. For a proportion, the standard deviation is the square root of p(1−p), which is largest near 0.5.

nₕ ∝ Nₕ × √(pₕ(1 − pₕ))

Medium and large businesses have proportions closer to 0.5, so they are more variable and get a slightly larger share.

Step 3. Compare the two

SE under proportional allocation = 1.831 points
SE under Neyman allocation = 1.824 points

Neyman is 0.8% more efficient. That is almost nothing, and saying so is more useful than presenting Neyman as a technique worth the trouble. It pays off when strata differ sharply in variability, and here they do not differ much.

Step 4. What stratification bought at all

A simple random sample of the same 600 gives an SE of 1.915 points. Proportional stratification gives 1.831, an improvement of 4.4%.

Modest, and typical. Stratification pays when the strata differ substantially on the outcome. Here they differ moderately, so the gain is real but small. Its other benefit is often more valuable: it guarantees you will have enough large businesses to report on separately, which a simple random sample does not.

Worked example two

What clustering costs you

Cluster sampling saves money by concentrating fieldwork. It is paid for in precision, and the bill is larger than most people expect.

Take the same 600 respondents, now reached by sampling whole workplaces.

Per clusterICCDesign effectEffective sample
100.051.45414
250.052.20273
100.202.80214
250.205.80103

design effect = 1 + (m − 1) × ICC, and effective n = 600 / design effect

Reading the bottom row

Twenty-five people from each of 24 workplaces, with an intracluster correlation of 0.20, gives 600 interviews worth about 103 independent observations. You collected six times more data than you can use.

The lever that matters

Look at the first and third rows. The ICC quadrupled and the design effect nearly doubled. Now compare rows one and two: holding ICC fixed and raising cluster size from 10 to 25 also inflates it sharply. More clusters with fewer people in each is almost always better than fewer clusters with more, for the same total sample and the same cost per interview.

Inflate the sample size at the design stage

If a power calculation says 400 and you intend to cluster, you need 400 multiplied by the design effect. At an ICC of 0.05 with 10 per cluster that is 580. Discovering this after fieldwork is not recoverable, which makes the design effect one of the few calculations that must be done first.

The same arithmetic governs the analysis, not just the design. Clustered data analysed as if independent produces standard errors that are too small, which is covered in the guide to clustered data and the unit of analysis.

Decision

Choosing between them

SituationChoose
Complete frame, population reachable remotelySimple random, or systematic if the list is long and unordered
Subgroups differ on the outcome and must be reported separatelyStratified, proportional unless variability differs sharply
Population is geographically dispersed and visits are costlyCluster, with many small clusters and an inflated sample
No frame exists and the population is hiddenSnowball, reported as non-probability
Qualitative study seeking depthPurposive, chosen for information richness
Fixed budget, demographic balance required, no frameQuota, reported honestly as non-probability

Most real studies combine them. A national survey might stratify regions, cluster within them, then sample systematically from each cluster. That is multistage sampling, and its variance calculation has to account for every stage.

What non-response does to a good design

A well-chosen sample is still only a plan. Response rates almost never come back flat across strata, and the differential is what causes the damage.

Return to the business survey. Suppose the response rates differ by size, as they usually do.

StratumSampledRespondedRateDesign weight
Small36019855%1.127
Medium18012670%0.886
Large604880%0.775

372 responses from 600. Now compute the headline figure two ways.

unweighted across respondents = 0.3469
weighted back to population shares = 0.3270

Two percentage points, in the direction of the group that answered most readily. Large businesses were 10% of the population but 13% of the respondents, and they had the highest value on the outcome, so the raw figure drifts upward. The weights above simply restore each stratum to its real share of the population.

Weighting fixes the composition, not the bias

Weighting repairs who is in your data. It cannot repair the possibility that non-responders within a stratum differ from responders in the same stratum. A small business that ignored the survey because it is struggling is not represented by a small business that replied. Report the response rate by stratum, state that you weighted, and do not claim the weighting removed non-response bias. The only real defence is a high response rate, which is a fieldwork problem rather than a statistical one.

Doing it

In R, SPSS and Stata

R

Drawing the sample, then analysing it with the design respected.

set.seed(2026)
srs <- frame[sample(nrow(frame), 600), ]
strat <- frame %>% group_by(size) %>% slice_sample(n = c(360, 180, 60)[cur_group_id()])

The analysis is where it matters. A stratified or clustered sample analysed with ordinary commands gives wrong standard errors.

library(survey)
d <- svydesign(ids = ~workplace, strata = ~size, weights = ~wt, data = samp)
svymean(~outcome, d)

ids declares the clusters and strata the strata. Omit ids on clustered data and the intervals come out far too narrow.

SPSS

Drawing a simple random sample is built in.

SAMPLE 600 FROM 10000.

For a stratified draw, sort and use SELECT IF within strata, or the Complex Samples module's sampling wizard. Analysis of a complex design needs that module: CSPLAN ANALYSIS to define the design, then CSDESCRIPTIVES or CSGLM. Ordinary FREQUENCIES on a clustered sample will understate the uncertainty.

Stata

Stata handles survey designs best of the three.

sample 6
svyset workplace [pweight = wt], strata(size) fpc(Nh)
svy: mean outcome
svy: proportion outcome, over(region)

svyset is declared once and every subsequent svy: command respects it. The fpc() option applies the finite population correction, which matters when you have sampled a large fraction of a small population and will otherwise overstate your uncertainty.

Record the seed

Any random draw should be reproducible. Set a seed, write it in the methods, and the exact sample can be regenerated if anyone asks how those particular units were chosen. It takes one line and it answers a question that is awkward to answer any other way.

Avoid these

Six mistakes that cost marks

MistakeWhy it matters
Calling a convenience sample random"Randomly approached shoppers" is convenience sampling. Random has a precise meaning and examiners check it.
Ignoring the design in the analysisA clustered sample analysed as independent gives intervals that can be half the correct width.
Not inflating for the design effect400 required with clustering at ICC 0.05 and 10 per cluster means recruiting 580. This cannot be fixed afterwards.
Systematic sampling into a cycleEvery 10th shift on a weekly rota selects the same weekday every time.
Treating quota as probabilityMatching on age and sex controls age and sex. Nothing else.
Describing the frame as the populationConclusions apply to the list you sampled from. Name the gap between it and the population you meant.

The second one is the most expensive, because it is invisible. Nothing in the output warns you. A clustered sample analysed with ordinary commands produces a p value, a confidence interval and a tidy table, all of which look exactly like correct output and are simply too narrow. Findings reported that way have been retracted.

The third is the most final. Every other item on this list can be addressed in the write-up by describing the limitation honestly. An under-powered clustered study cannot, because the data does not exist and fieldwork has ended.

Clustering, and not sure what to recruit?

A power calculation returning 400 means recruiting 580 once clustering is accounted for, and that is not something you can fix after fieldwork. Send the design and you get the target, the allocation, and an analysis plan that matches how you sampled.

See power analysis and sample size
Elaine HalliburtonElaine HalliburtonProfessor of Applied Statistics

Writing it up

How to report it

✓The target population, the frame, and the difference between them.
✓The method, named precisely, and why it was chosen over the alternatives.
✓For stratified designs, the strata, the allocation and the rule used.
✓For clustered designs, the number of clusters, the average size and the assumed or observed ICC.
✓The achieved sample against the target, with reasons for any shortfall.
✓Whether the analysis accounted for the design, and in which software.

For non-probability designs add one sentence placing the limit on your own claims, for example: "Participants were recruited purposively for maximum variation. Findings describe this sample and are not generalisable to the wider population."

Common questions

Frequently asked questions

What is the difference between probability and non-probability sampling?

Probability sampling gives every member of the population a known, non-zero chance of selection, which is what allows a confidence interval to describe how far your estimate might be from the truth. Non-probability sampling does not, so the interval still computes but no longer means what it appears to mean.

What is stratified sampling?

Dividing the population into groups that differ on the outcome, then sampling within each. It guarantees representation of every group and usually improves precision, though often modestly: in a worked example it reduced the standard error by 4.4% against a simple random sample of the same size.

Should I use proportional or Neyman allocation?

Proportional unless the strata differ sharply in how variable the outcome is. Neyman samples more heavily where there is more variation, but when variability is similar across strata the gain is tiny; in the worked example it was 0.8%, which rarely justifies the extra complexity.

What is purposive sampling?

Deliberately selecting participants for a characteristic rather than at random, usually to obtain information-rich cases. It is the standard and correct approach in qualitative research, where maximum variation or extreme cases tell you more than a random draw. It does not support generalisation to a wider population and should not claim to.

Why does cluster sampling need a larger sample?

People within a cluster resemble each other, so each additional person in the same cluster adds less new information. The design effect is 1 + (m − 1) × ICC. With 25 per cluster and an ICC of 0.20 the design effect is 5.80, so 600 interviews are worth about 103 independent observations.

Is snowball sampling acceptable in academic research?

Yes, where no sampling frame exists, which is common for hidden or stigmatised populations. It is a recognised method with a substantial literature. Report it as non-probability, describe how the chains started and how long they ran, and do not attach inferential claims about the wider population.

How large should my sample be?

That is a power question rather than a sampling one, and it depends on the effect you need to detect and the precision you need. Whatever the calculation returns, inflate it for the design effect if you are clustering, and again for expected non-response.