Back
Sampling Methods: Stratified, Cluster, Random and Quota
Matthieu Saussaye
Sampling method is how you choose who gets asked, and it decides what your results are allowed to mean. The first split is between probability sampling, where every member of the population has a known, non-zero chance of selection, and non-probability sampling, where they do not. Only probability sampling supports a defensible margin of error.
The short version:
Simple random sampling draws names at random from a complete list. Clean, unbiased, and often impossible because the list does not exist.
Systematic sampling takes every nth unit from an ordered list. Easier to execute, breaks if the list has a hidden cycle.
Stratified sampling divides the population into groups first, then samples within each. More precise than simple random when the groups genuinely differ.
Cluster sampling samples whole groups such as schools or districts, then surveys within them. Much cheaper in the field, less precise for the same sample size.
Quota sampling fills preset counts per subgroup without random selection. Fast and controllable, non-probability, and the standard for online panels.
Convenience and snowball sampling take whoever is reachable, or whoever a respondent refers. Useful for exploration and hard-to-reach groups, not for estimating population figures.
Sampling methods compared
Every method buys something and charges for it somewhere else, usually in cost, precision or the strength of the claim you can make afterwards.
Method | How it works | Best for | Main risk |
|---|---|---|---|
Simple random | Every unit on a complete list has an equal chance of selection | Populations with a real, complete sampling frame | The frame rarely exists, and small subgroups can be missed by chance |
Systematic | Pick a random start, then take every nth unit from an ordered list | Long lists and physical fieldwork where randomization is awkward | Periodicity: if the list cycles in step with your interval, the sample is biased |
Stratified | Split the population into strata, then sample randomly inside each one | When subgroups differ and you need reliable subgroup estimates | Needs known strata membership up front, and weighting if allocation is not proportional |
Cluster | Randomly select whole groups, then survey all or some members within them | Geographically dispersed populations with no individual-level list | Lower precision per respondent, because people inside a cluster resemble each other |
Quota | Fill fixed counts per subgroup, recruiting non-randomly until each is full | Online panels, fast commercial studies, controlled sample shape | Non-probability: who fills each quota is self-selected, so no true margin of error |
Convenience | Survey whoever is easiest to reach | Pilots, question testing, early exploration | Unknown and usually large bias; cannot support population estimates |
Snowball | Existing respondents refer further respondents | Rare or hard-to-reach populations with no list | Sample follows social networks, so it over-represents well-connected people |
The first four rows are probability methods. The last three are not, and the difference is not a technicality: it is what separates "we estimate 42%, plus or minus 3 points" from "42% of the people we reached said this".
Probability versus non-probability sampling
Probability sampling means every unit in the population has a known, non-zero probability of being selected, and that selection is made by a random mechanism you control. Statistics Canada, Survey Methods and Practices, 2010 states the two criteria: "one is that the units be randomly selected, the second is that all units in the survey population have a non-zero inclusion probability in the sample and that these probabilities can be calculated". That property is what allows sampling error to be calculated. A margin of error is a statement about how much results would vary across repeated random samples. Without random selection, the calculation has nothing to describe.
Non-probability sampling means selection depends on something other than chance: who was available, who volunteered, who a friend referred, who was still in the panel and willing to answer at 9pm. The resulting sample can still be informative and it is what most commercial research actually uses. What it cannot do is support the claim that the difference between your sample and the population is only sampling noise. AAPOR has long maintained that "reporting margin of sampling error with opt-in or self-identified samples is misleading" (Baker et al., AAPOR Task Force on Non-Probability Sampling, 2013).
Two practical notes that cut through the purity debate:
In practice, almost nothing is purely probability-based any more. Nonresponse means that even a perfect random draw ends up with a self-selected set of people who agreed to answer. That is one reason response rates matter so much: a low response rate erodes the very property that made the design probabilistic.
Non-probability is not the same as worthless. A well-managed quota sample from a large panel, weighted to known population figures, is the normal basis for commercial market research. As the AAPOR task force (Baker et al., 2013) concluded, non-probability samples "may be appropriate for making statistical inferences, but the validity of the inferences rests on the appropriateness of the assumptions underlying the model and how deviations from those assumptions affect the specific estimates". The honest move is to describe the method rather than to dress it up with a margin of error it did not earn.
Simple random and systematic sampling
Simple random
Take a complete list of the population, the sampling frame, and draw from it at random. Every unit has the same probability of selection. It is the reference design, the one all other methods are compared against, and it is unbiased by construction.
Its practical problem is the frame. A complete, current list of "all people in a country" or "all buyers in a category" usually does not exist. Where a frame does exist, such as your own customer database or an employee directory, simple random sampling is straightforward and usually the right default.
It has one statistical weakness worth knowing. Random draws can, by chance, under-represent a small subgroup. If you need reliable numbers for a group that is 4% of the population, simple random sampling will give you few of them and no guarantee of how few.
Systematic
Order the frame, pick a random starting point, then take every nth unit. With 10,000 records and a target of 500, you take every 20th. It is easier to execute than drawing 500 random numbers, and it spreads the sample evenly across the list.
The failure mode is periodicity. If the list has a repeating pattern whose cycle aligns with your interval, the sample systematically hits the same kind of unit every time. A list of store visits ordered by day, sampled every 7th, can end up entirely on Mondays. Statistics Canada (2010) uses the same example for traffic flow and adds the awkward part: "in most cases, periodicity is not known in advance". Check the ordering before you trust the design; if the order is meaningful, randomize it first.
When stratified sampling beats simple random
Stratified sampling divides the population into mutually exclusive strata, such as region, company size or age band, and then draws a random sample within each stratum. Every unit still has a known probability of selection, so it remains a probability method.
It beats simple random sampling in two situations.
You need subgroup estimates
If the report has to say something about each region separately, a simple random sample gives you whatever count each region happens to receive, which may be too few to say anything. A stratified sample lets you set the count per stratum in advance. Over-sampling a small but important stratum is standard practice, and the price is that you must weight the results back to true population proportions before reporting a total. Skip the weighting and your national figure is quietly the average of an artificial population.
The strata differ on what you are measuring
This is the precision argument, and it is the one people miss. When variation within each stratum is small and variation between strata is large, stratification removes the between-stratum variability from your sampling error. For a fixed sample size, a stratified estimate of the overall mean is then more precise than a simple random one. Statistics Canada (2010) states the condition directly: to beat simple random sampling "there must be strong homogeneity within a stratum ... and the strata themselves must be as different as possible", which happens "if the stratification variables are correlated with the survey variable of interest".
The corollary matters just as much: if the strata do not differ on your outcome, stratifying by them buys almost nothing. Stratifying a customer survey by first initial is valid and pointless. Choose strata that you have reason to believe relate to the thing you are measuring.
The requirement is that you must know each unit's stratum before sampling. That is easy for a customer database with a region field and impossible for a characteristic you can only learn by asking. When the classifier is only knowable from the survey itself, you are usually doing quota sampling, not stratified sampling.
How cluster sampling trades precision for fieldwork cost
Cluster sampling reverses the logic of stratification. Instead of sampling within every group, you randomly select whole groups and then interview inside the selected ones. Sample 30 schools out of 900, then survey pupils in those 30. In one-stage cluster sampling you take everyone in the chosen clusters; in two-stage sampling you sample within them as well.
The reason to do this is almost always cost. If interviewing means physically travelling, concentrating your fieldwork in 30 locations rather than spreading it across 900 changes the budget by an order of magnitude. Cluster sampling also works when no list of individuals exists but a list of clusters does. There is no register of every household, but there is a list of districts.
The cost is precision. People inside a cluster tend to resemble each other, and that similarity is measured by the intraclass correlation. Because they are alike, each additional respondent from an already-sampled cluster adds less new information than an independent respondent would. The result is a design effect, which Statistics Canada (2010) defines as "the ratio of the sampling variance of an estimator under a given design to the sampling variance of an estimator under SRS of the same sample size". The same source notes that cluster sampling "is usually a less statistically efficient sampling strategy than SRS" and is chosen because "sampling clusters can greatly reduce the cost of collection, particularly if the population is spread out and personal interviews are conducted". Your effective sample size ends up smaller than your raw count.
Two implications follow directly:
Many small clusters beat a few large ones. Adding clusters buys more precision than adding respondents inside existing clusters. Statistics Canada (2010): "When neighbouring units are similar, it is more statistically efficient to select many small clusters rather than a few, larger clusters."
The analysis must account for the design. Standard errors computed as if the sample were simple random are too small, which makes differences look significant when they are not. Use survey-design-aware estimation.
Stratified and cluster sampling are often confused because both involve groups. The distinction is simple: stratification samples from every group and increases precision, clustering samples only some groups and reduces cost.
Quota sampling, and how it works in practice
Quota sampling sets target counts for subgroups, then recruits until each target is filled. A study might specify 500 respondents split by age, gender and region to match national proportions. Once the 25-34 cell is full, later respondents in that bracket are screened out.
Structurally it resembles stratified sampling: the population is divided into cells with a planned count each. The difference is decisive. In stratified sampling, units within each stratum are selected at random from a frame. In quota sampling they are not; whoever shows up and qualifies gets in. That makes it non-probability, no matter how well the margins match the census. Statistics Canada (2010) draws the same line: "In probability sampling, the units are selected randomly while in quota sampling a non-random method is used", and "contacted units that are unwilling to participate are simply replaced by units that are, in effect ignoring nonresponse bias".
Why it dominates commercial research anyway:
It is fast and predictable. You know the shape of your final sample before fieldwork starts.
It needs no frame. Only a screener that identifies which cell a respondent belongs to.
It guarantees subgroup counts. The segments you promised to report on will have enough respondents to report.
What it hides is who filled each cell. Matching the population on age and gender does not make the sample representative on anything else. The people who complete online surveys for a small incentive differ from those who do not in ways that quotas on demographics never touch: available time, interest in the topic, and how many surveys they take per week. Quota control on demographics is a floor, not a guarantee. If you are buying sample from a panel, the composition and management of that panel matters as much as your quota grid, which is the subject of panel research.
Interlocked versus non-interlocked quotas
A non-interlocked quota grid sets independent targets: 50% women, 20% aged 18-24. Each margin comes out right while the combination can be badly skewed, for example if nearly all of your 18-24s are men. Interlocked quotas specify the cells themselves, such as 10% women aged 18-24. Interlocking gives better control and is harder to fill, because rare cells become the bottleneck at the end of fieldwork. Interlock the variables that interact with your outcome, and leave the rest as margins.
Convenience and snowball sampling
Convenience sampling means surveying whoever is easiest to reach: your email list, your social followers, people walking past. The bias is unknown in size and direction, and it is usually large. Your followers are not your market, they are the part of your market that already likes you.
That does not make it useless. It is the right method for testing whether a questionnaire is comprehensible, for early exploratory work where you are trying to find out what the questions should be, and for problem discovery before you commit budget. The rule is that findings are hypotheses, not estimates. Never attach a percentage to a population from a convenience sample.
Snowball sampling recruits through referral: each respondent nominates others. It exists to reach populations with no list and low incidence, such as people with a rare condition or holders of a specialized job title. It works where nothing else does, and it inherits the shape of the social network it travels through, over-representing the well-connected and clustering within communities. Report it as what it is: a sample drawn through referral chains, described rather than projected.
Sample size is a separate decision
Sample size and sampling method are two different questions, and confusing them is common. Method determines whether your sample is systematically off. Size determines how much random noise sits around your estimate.
A large sample drawn badly is not better than a small sample drawn well. Increasing the size of a biased sample makes the estimate more precisely wrong: the confidence interval narrows around a number that was never centered on the truth. Deciding on the method comes first.
How your sampling choice constrains what you can claim
Every sampling design licenses a specific sentence in your report. Writing a stronger sentence than the design supports is the most common way research becomes misleading without anyone lying.
Probability sample with good response: "We estimate 42% of the population, with a margin of error of plus or minus 3 points at 95% confidence."
Stratified probability sample: the same claim, plus reliable subgroup statements, provided results are weighted back to population proportions.
Cluster sample: the same claim, but with standard errors computed for the design. The effective sample size is smaller than the number of interviews.
Quota sample from a panel: "Among 1,000 panel respondents quota-controlled to match national age, gender and region proportions, 42% said X." Describe the method; do not present a sampling margin of error as if selection had been random.
Convenience sample: "Among 180 of our own customers who responded, 42% said X." No projection beyond the responders.
Snowball sample: "Among 60 respondents recruited through referral, most described X." Qualitative framing only.
Weighting does not convert one into another
Weighting adjusts your sample so its composition matches known population figures on the variables you weight by. It is standard, it is useful, and it corrects real imbalances. It does not turn a non-probability sample into a probability one. Weighting can only fix bias on the variables you weight and on anything closely related to them. It cannot fix a difference between responders and non-responders that is unrelated to age, gender and region. Heavy weights also inflate variance, which reduces effective sample size, so a heavily weighted sample is less precise than its raw count suggests.
Report the design, always
A methods note should state the population you targeted, the frame or source of respondents, the selection method, the quotas or strata used, the field dates, the achieved sample size, and any weighting applied. If a reader cannot reconstruct how people got into your sample, they cannot judge the numbers. This matters most when the design was not probabilistic: Baker et al. (2013) hold that "whenever non-probability sampling methods are used, there is a higher burden than that carried by probability samples to describe the methods used to draw the sample, collect the data, and make inferences". This is also what separates a serious study from a chart with a percentage on it, and it is worth checking before you commission work from an external research agency.
Picking a method: a short decision path
Do you have a complete list of the population? If yes, simple random or systematic sampling is available. If no, skip to quota or cluster designs.
Do you need reliable numbers for specific subgroups? If yes, stratify on those subgroups, or set quotas for them, and plan to weight.
Is fieldwork geographically expensive? If yes, cluster, and accept the design effect in your sample size calculation.
Is the population rare with no list? Snowball or targeted recruitment, reported descriptively.
Are you exploring rather than estimating? Convenience is fine. Say so in the report.
The method that fits is generally the most rigorous one your frame and budget actually allow. The judgment then goes into describing it accurately, which is the same discipline that separates qualitative from quantitative claims.
References
Baker, R., Brick, J. M., Bates, N. A., Battaglia, M., Couper, M. P., Dever, J. A., Gile, K. J., and Tourangeau, R. Report of the AAPOR Task Force on Non-Probability Sampling. American Association for Public Opinion Research, June 2013.
Statistics Canada. Survey Methods and Practices, Catalogue no. 12-587-X. Statistics Canada, 2010.
Get the sample shape right, and the depth too
Sampling decides who answers. It says nothing about how much you learn from each of them. Most studies solve the first problem carefully and then collect two clicks and a one-word comment from every respondent they worked so hard to recruit.
SmartInterview handles quota management so cells fill the way you specified, and runs the interview itself by voice or text with an AI that follows up on vague answers in the respondent's own language. Open responses come back coded into themes, so the qualitative depth is countable next to your quota-controlled numbers.
Design your next study end to end: start free or talk to our team.
Frequently Asked Questions
What is the difference between stratified and cluster sampling?
Both divide the population into groups, but they use them in opposite ways. Stratified sampling samples from every group, which increases precision when the groups differ on what you are measuring. Cluster sampling randomly selects a subset of groups and surveys only inside those, which cuts fieldwork cost but lowers precision because people within a cluster resemble each other.
What is a stratified sample?
A stratified sample is drawn by first dividing the population into mutually exclusive strata, such as region or company size, and then taking a random sample within each stratum. It requires knowing each unit's stratum before sampling. It guarantees enough respondents in each subgroup and, when the strata differ on the outcome, produces a more precise overall estimate than simple random sampling of the same size.
When should you use cluster sampling?
Use it when the population is geographically dispersed, fieldwork requires travelling, and no list of individuals exists but a list of clusters does. Concentrating interviews in a limited number of randomly selected locations reduces cost sharply. Budget for the design effect: because responses within a cluster are correlated, you need a larger sample than a simple random design to reach the same precision.
Is quota sampling the same as stratified sampling?
No. They look similar because both set target counts per subgroup, but stratified sampling selects units randomly within each stratum from a sampling frame, while quota sampling accepts whoever qualifies until each cell is full. That makes stratified sampling a probability method with a calculable margin of error, and quota sampling a non-probability method without one.
What is the difference between probability and non-probability sampling?
In probability sampling, every member of the population has a known, non-zero chance of selection and the selection is random, which is what makes sampling error calculable. In non-probability sampling, selection depends on availability, self-selection or referral. Non-probability samples can still be informative, but you cannot claim that the gap between your sample and the population is only random noise.
Does weighting fix a biased sample?
Only partly. Weighting corrects imbalance on the variables you weight by, and on characteristics closely related to them. It cannot correct differences between responders and non-responders that are unrelated to those variables, and it does not convert a non-probability sample into a probability one. Heavy weights also increase variance, reducing the effective sample size below the number of interviews you actually collected.
