Sampling Methods: Probability and Non-Probability Types Explained

A sampling method is the defined procedure a study uses to select which members of a population are actually observed, measured, or surveyed. Sampling methods split into two families: probability sampling, where every population member has a known, non-zero chance of selection, and non-probability sampling, where selection is not governed by a defined chance mechanism. That distinction is the single most consequential choice a Methods section reports, because it determines whether standard inferential statistics and generalizability claims back to a population are justified, or whether the sample can only be defended on other grounds (depth, access, theoretical relevance).

Probability sampling methods

In probability sampling, every unit in the population has a known, non-zero probability of being selected, typically via a random mechanism. This is what allows a sample statistic to be used to estimate a population parameter with a quantifiable margin of error — the basis for confidence intervals and most significance testing. See CASRAI’s comparison of population vs. sample for how parameters and statistics relate.

• Simple random sampling — every member of the population has an equal, independent chance of selection, typically implemented by drawing from a complete roster using a random-number generator. It is the baseline method most statistical inference formulas assume, but it requires a complete sampling frame (a full list of the population) and, with small samples, can by chance still produce meaningfully unequal representation of subgroups.
• Stratified sampling — the population is first divided into relevant, non-overlapping subgroups (strata) — by institution type, career stage, disease severity, geographic region — and a random sample is drawn independently from within each stratum, usually in proportion to the stratum’s size in the population. This guarantees representation of subgroups that simple random sampling might under-sample by chance alone, and improves precision when the stratifying variable is meaningfully related to the outcome. It requires knowing each population member’s stratum in advance, which is not always available.
• Systematic sampling — a starting point is chosen at random from an ordered list, then every k-th unit after it is selected (e.g., every 10th name on a roster). It is simpler to execute than simple random sampling and spreads the sample evenly across the list, but it can introduce bias if the list has a hidden periodic pattern that lines up with the sampling interval (e.g., a roster ordered by shift schedule where every k-th person shares a shift).
• Cluster sampling — the population is divided into naturally occurring groups (clusters) — clinics, classrooms, geographic areas — and a random sample of whole clusters is selected, with every member of a selected cluster included (or a further random sample drawn within each selected cluster, in multi-stage cluster sampling). It is far cheaper and more practical than sampling individuals scattered across an entire population, especially when no complete individual-level sampling frame exists, but it typically yields less statistical precision than simple random sampling for the same total sample size, because members of the same cluster tend to resemble each other more than the population at large (a design effect that analysis needs to account for).

Non-probability sampling methods

In non-probability sampling, selection is not governed by a known chance mechanism, so the probability that any given population member ends up in the sample cannot be calculated. This makes standard inferential statistics and population-level generalizability claims difficult to justify without additional argument, but non-probability methods remain widely and legitimately used — particularly in qualitative research, exploratory or pilot work, and studies of populations with no accessible sampling frame.

• Convenience sampling — participants or items are included because they are readily accessible to the researcher (students in the researcher’s own course, patients presenting at a single clinic, volunteers responding to a call). It is the fastest and cheapest method to execute, which makes it common in pilot studies and early-stage research, but it carries the highest risk of an unrepresentative sample among the non-probability methods, since who is “convenient” often correlates with the outcome being studied.
• Purposive (judgmental) sampling — the researcher deliberately selects participants believed to be information-rich or particularly relevant to the research question, based on defined selection criteria and the researcher’s judgment. It is standard practice in qualitative research designed to explore a phenomenon in depth rather than estimate its prevalence, and includes named variants such as maximum-variation sampling (deliberately selecting cases that differ widely on key characteristics) and typical-case sampling (selecting cases judged representative of the “normal” or average case).
• Snowball sampling — existing participants recruit or refer additional participants from their own networks, expanding the sample outward in successive waves. It is the standard approach for reaching populations with no accessible sampling frame and no public list of members — hidden, stigmatized, or hard-to-identify populations (for example, people who share a rare experience or an illegal or highly private behavior) — but it risks over-sampling within a single social network and under-representing anyone outside the recruitment chains that happen to form.
• Quota sampling — the researcher sets target quotas for key subgroups (e.g., a fixed number of participants by age band and sex) meant to mirror their known proportions in the population, then fills each quota using convenience or judgment rather than random selection within the subgroup. It resembles stratified sampling in structure — both divide the population into subgroups first — but differs in the critical respect that selection within each subgroup is non-random, so quota sampling does not carry stratified sampling’s statistical guarantees, only a closer surface resemblance to the population’s demographic composition.

Choosing a sampling method: the core tradeoff

The central tradeoff is representativeness and generalizability versus feasibility, cost, and access. Probability methods are the only methods that support formal statistical inference back to a defined population with a quantifiable margin of error, which is why they are expected in most quantitative survey research, epidemiology, and any study making a explicit population-level claim. Non-probability methods trade that formal guarantee for speed, lower cost, and access to populations or phenomena a probability design cannot reach at all — they are the norm, not a compromise, in most qualitative research, where the goal is depth of understanding of a phenomenon rather than a precise population estimate.

A larger sample does not fix a non-representative sampling method: representativeness is a property of how the sample was selected, not of how many units it contains. A study should report its sampling method explicitly enough that a reader can judge what population, if any, the findings are entitled to generalize to — this is the same principle covered in more depth in CASRAI’s comparison of population vs. sample, which also covers sampling bias and the parameter/statistic distinction.

Sampling vs. randomization: a common conflation

Sampling and randomization are frequently confused but answer different design questions. Sampling concerns how participants are drawn from a population into a study in the first place; randomization concerns how participants already enrolled in a study are subsequently allocated between comparison arms or conditions. A trial can randomize its arms while still drawing its overall sample by convenience, and a survey can use rigorous probability sampling with no randomization at all, because it has no comparison arms to allocate. See CASRAI’s dictionary entry on randomization methods in clinical trials for the allocation side of this distinction.

Reporting sampling method in a manuscript

A Methods section should name the specific sampling method used (not just “a sample was recruited”), describe the sampling frame or recruitment mechanism, and, for probability methods, report how any stratification or clustering variables were defined. For non-probability methods, it should state explicitly that population-level generalizability is limited and, where relevant, describe what steps (e.g., maximum-variation purposive sampling, quota targets) were taken to broaden the sample’s range despite the lack of random selection.

Illustrative example, not a real study: a researcher studying research-data-management practices across a university might use stratified random sampling — dividing faculty into strata by discipline and career stage, then randomly sampling within each stratum — to support a claim about typical practice across the whole institution. A separate qualitative interview study of the same topic might instead use purposive sampling to deliberately recruit faculty known to have unusually mature or unusually absent data-management practices, in order to explore the range of practice in depth rather than estimate its prevalence; that qualitative sample would not be presented as representative of the university as a whole.

Related terms

See CASRAI’s comparison of population vs. sample for the broader concepts of sampling frame, sampling bias, parameters, and statistics that sampling method choice feeds into, and randomization methods in clinical trials for how allocation to study arms differs from sampling into the study. CASRAI’s dictionary entry on research questionnaire covers a common instrument fielded to a sample once it has been selected, and the entry on confidence interval covers how a probability sample’s estimate is reported with a quantified margin of error.