By the Wavefield Research team · Published Oct 12, 2026
A sampling method is the rule that decides who gets into your survey. There are nine worth knowing, in two families: probability methods, where chance does the selecting, and non-probability methods, where the researcher or the recruitment channel does. The choice sets what your results can honestly claim.
Most guides to sampling methods in statistics stop at textbook definitions. This one adds the part practitioners need: which methods online research actually runs on, what that does to the margin of error, and how weighting repairs — partially — what the sampling method cannot deliver.
Probability sampling methods give every population member a known chance of selection, which is what makes a formal margin of error possible. Non-probability sampling methods trade that guarantee for speed, cost, and reachability. Neither family is wrong — but reports go wrong when a study run with one family borrows the statistical claims of the other.
Probability
Simple random sampling gives every member of the population an equal chance of selection — draw names from the full list at random, nothing more clever than that.
Example: Pick 400 employees at random from a complete HR roster of 4,000 for a culture survey.
The textbook ideal, and the baseline every margin-of-error formula assumes. Needs a complete list of the population (a sampling frame), which is exactly what most commercial studies do not have.
Probability
Systematic sampling takes every kth record from an ordered list after a random starting point — every 10th customer in the billing file, say.
Example: From 20,000 subscribers, survey every 50th account starting at a randomly chosen row.
Operationally simpler than a true random draw and usually just as good. The one trap: a list sorted in a cycle that matches your interval (every 7th record is a weekend signup) quietly biases the draw.
Probability
Stratified sampling splits the population into groups (strata) first — region, plan tier, age band — then draws randomly inside each group.
Example: Split customers into free, standard, and enterprise tiers, then randomly sample each tier so even the small enterprise group yields enough completes to analyze.
Guarantees subgroup coverage and tightens precision compared with a simple random draw of the same size. The deep dive below covers proportionate vs. disproportionate allocation.
Probability
Cluster sampling randomly selects whole groups — schools, stores, city blocks — and then surveys everyone (or a random subset) inside the chosen clusters.
Example: Randomly pick 30 of a chain's 600 stores and survey the staff at those 30, instead of a few people at all 600.
The economics play: travel and recruitment costs collapse when respondents come in bundles. The price is wider error margins, because people in a cluster resemble each other.
Probability
Multistage sampling chains the methods above: randomly select regions, then households within regions, then one adult within each household.
Example: National face-to-face studies and the big government surveys — select counties, then blocks, then households, then a person.
How official statistics are actually collected. Rarely justified for commercial work — if you are not knocking on doors, you rarely need it.
Non-probability
Convenience sampling takes whoever is easiest to reach — a link on your homepage, an intercept outside one store, your own social followers.
Example: A pop-up survey shown to whoever happens to visit the website this week.
Cheap, fast, and skewed by construction: you learn about the people you can reach, who are rarely the people you need. Fine for pilots and question testing, dangerous for decisions.
Non-probability
Quota sampling recruits until each predefined cell hits its target — 150 men and 150 women, spread across age bands — then closes the cell and screens further matches out.
Example: A 300-complete consumer study where panel respondents flow in until age × gender cells match census shares.
The workhorse of commercial research and the method behind most online panel studies. How the cells get set, tracked, and closed in practice is its own guide.
Non-probability
Purposive (or purposeful, or judgment) sampling hand-picks respondents who fit a profile — power users, recent churners, category experts — because their specific experience is the point.
Example: Interviewing 25 customers who canceled in the last 60 days to understand churn.
The standard for qualitative work, where saturation matters more than representativeness. No one claims the 25 churners stand in for all customers — and no one needs them to.
Non-probability
Snowball sampling asks each respondent to refer the next — the sample grows through the population's own network.
Example: Reaching freelance set designers by asking each interviewee to introduce two more.
Sometimes the only way to reach rare or hidden populations. Network structure drives who appears, so treat the results as exploration, not measurement.
Quota sampling gets the full treatment in its own guide — quota sampling vs. random sampling — including how quota cells are set, tracked, and closed during a live field.
Stratified sampling earns its popularity by solving two problems at once. First, coverage: a simple random draw of 400 customers can, by bad luck, land only 12 enterprise accounts — too few to analyze. Stratifying by tier makes the enterprise cell's size a design decision instead of a coin flip. Second, precision: when the strata differ from each other and are internally similar, stratified estimates carry less variance than a simple random sample of the same size. You get a tighter answer for the same budget.
Proportionate allocation sizes each stratum's sample to match its population share: if enterprise is 8 percent of customers, it is 8 percent of the sample. Disproportionate allocation deliberately oversamples small-but-important strata — drawing 100 enterprise accounts instead of 32 — so the subgroup is analyzable on its own, then weights each stratum back to its true share for the combined readout. The oversample is not a flaw to apologize for; it is the standard way to buy subgroup precision without pretending the subgroup is bigger than it is.
In market research practice, stratified sampling mostly appears in list-based studies — employee surveys from an HR roster, customer studies from a CRM export — because those are the rare cases where a complete frame exists. Field it by splitting the list, drawing randomly within each stratum, and inviting by email or SMS. When there is no list and respondents come from a panel, what looks like stratification is almost always quota sampling — the distinction the FAQ below pins down.
Here is the sentence most sampling guides avoid: nearly all commercial online research — brand trackers, concept tests, most published polls — runs on non-probability samples from opt-in panels, managed with quota sampling. Nobody has a list of all consumers to draw from randomly. The respondents are people who signed up to take surveys, and the method's job is to shape that flow into something that mirrors the population on the dimensions that matter.
This is not a scandal; it is a documented trade-off. The American Association for Public Opinion Research examined the practice in its Task Force Report on Non-Probability Sampling and its conclusion is the useful one: non-probability samples are fit for many purposes, but their quality depends on how the sample is controlled and adjusted, and they should never be described as if chance had selected them. The probability alternative still exists — Pew Research Center's American Trends Panel recruits members offline from random address-based samples — but that machinery costs more than most study budgets, which is precisely why the panel-plus-quotas model won the commercial market.
Done well, the model has real controls: quota cells on the variables that drive the answers, screening that verifies who respondents are, quality checks that catch speeders and straight-liners, and weighting that trues up the remaining skew. Done badly, it is a convenience sample with paperwork. The difference is visible in the field report, not the invoice.
Every sampling method hands the analyst a sample that misses the population somewhere — quota cells cover age and gender but not education; a stratified employee study over-represents the oversampled stratum by design. Post-stratification weighting is the repair step: compare the sample's shape to known population targets, then give each respondent a weight so the weighted sample matches those targets. With several dimensions at once, the standard algorithm is raking — iterate across the targets until they all hold — which is exactly what the survey weighting guide walks through.
Two limits keep the repair honest. Weighting can only fix the dimensions you weight on: a sample skewed on something unmeasured stays skewed. And weights are not free — spreading extra weight across few respondents inflates variance, which is why weighted results should report effective sample size rather than raw n. Choosing bracket sets that match your weighting targets starts at questionnaire time, not analysis time — the demographic questions guide treats option sets as weighting targets in disguise, which is what they are.
Sampling error is the random part: observe 400 people instead of everyone and your estimate wobbles around the truth, by an amount that shrinks predictably as n grows. That predictable wobble is what a margin of error measures, and the sample size calculator turns it into a planning number. Sampling bias is the systematic part: the selection rule itself tilts who gets in, and no increase in sample size untilts it.
The margin-of-error question follows directly. A formal margin of error assumes a probability sample; on a quota sample there is no random-selection mechanism for the math to describe. Serious practice handles this with a convention: report the margin a probability sample of the same effective size would carry, label the sampling method plainly, and let readers judge. What crosses the line is the convenience-sample press release quoting ±3 points as if the arithmetic blessed the recruitment. The honest phrasing is boring and correct: “results are quota-balanced to census targets; a probability sample of this size would carry a margin of about ±5 points.”
Have a complete list of the population? Use it — this is the probability family's home turf. Simple random or systematic sampling for a uniform population; stratified sampling when subgroups matter or vary in size. Employee and customer studies live here.
No list, need population estimates? Quota sampling on a research panel, cells set on the variables that drive your topic, weighting after fielding, method stated in the report. This is the standard commercial design, and it is defensible exactly as long as it is described truthfully.
Need depth from specific people, not estimates? Purposive sampling, recruited deliberately, analyzed as qualitative evidence. Saturation, not representativeness, is the stopping rule.
Rare or hidden population? Snowball sampling, reported as exploration. Official statistic or high-stakes public claim? Budget for a true probability design — multistage or address-based — and the field costs that come with it.
Lists that say five usually mean the probability family: simple random, systematic, stratified, cluster, and multistage sampling. The full map adds the non-probability family — convenience, quota, purposive, and snowball sampling — which is where most commercial online research actually lives.
Cluster sampling randomly picks whole groups and surveys inside the chosen groups only; stratified sampling divides the population into groups and samples inside every group. Stratification exists to tighten precision; clustering exists to cut cost, and it widens error because people within a cluster resemble each other.
Sampling bias is a systematic mismatch between who ends up in the sample and the population it claims to describe — website visitors standing in for all customers, or panel joiners standing in for all voters. More completes do not fix it: a bigger biased sample is just a more confident wrong answer.
Sampling error is the random wobble from observing a sample instead of everyone — it shrinks as sample size grows and is what a margin of error describes. Non-sampling error is everything else: bad question wording, nonresponse, careless answers, coverage gaps. It does not shrink with sample size, which is why questionnaire quality matters as much as n.
No, and the confusion is common because both use population groups. Stratified sampling draws randomly inside each stratum from a full list — it is a probability method. Quota sampling fills each cell with whoever qualifies first, usually from a panel — a non-probability method. Same cell math, different selection mechanics, different statistical guarantees.
For most commercial studies the honest answer is quota sampling on a panel, with weighting afterward and the method stated plainly in the report. A probability design is worth its cost when the stake is an official statistic or a high-stakes public claim. The wrong answer is a convenience sample dressed up with a margin of error.
Wavefield runs quota cells, quality checks, and raking weights as built-in features — describe your study and the AI drafts the questionnaire, screeners, and quota plan for you to review.
Start a study