Sampling is the process of selecting a subset of units, observations or measurements from a larger set for the purpose of obtaining information about that larger set or efficiently conducting an investigation. The units selected may be people, households, businesses, geographical locations, biological organisms, manufactured products, transactions, records or experimental measurements. Sampling is necessary when examining every member of a population would be impractical, expensive, destructive or impossible.
The validity of conclusions depends not simply on sample size but on how the sample was selected, what population the sampling procedure actually represents and whether the measurements obtained accurately capture the variables of interest. Sampling therefore involves both a selection mechanism and a relationship between the observed sample and the target population.
Probability sampling uses a defined random mechanism so that units have known or calculable probabilities of selection. Simple random sampling gives eligible units an equal selection probability under the relevant design, while systematic sampling selects units according to a specified interval after an appropriate starting procedure. Stratified sampling divides a population into defined strata and samples within them, often improving representation or precision when strata differ meaningfully.
Cluster sampling selects groups or clusters rather than individual units and can reduce fieldwork costs, although observations within clusters may be correlated. Multistage sampling combines several selection stages, such as choosing regions, then institutions, then individuals. These designs require appropriate weighting and variance estimation when probabilities of selection differ.
Non-probability sampling selects units without a probability mechanism that permits every population member's selection probability to be known. Convenience sampling uses readily accessible observations, while purposive or judgement sampling deliberately selects units considered relevant to a research objective. Quota sampling seeks specified numbers within categories but does not necessarily use random selection within those categories. Snowball sampling recruits additional participants through existing participants and can be useful for populations that are difficult to enumerate or reach. These approaches can be appropriate for exploratory research, qualitative investigation or specialised contexts, but ordinary statistical generalisation to a defined population may require assumptions that are not justified by the sampling procedure alone. A large non-probability sample can still be systematically unrepresentative.
Sampling error is the variation that arises because only part of a population is observed rather than the entire population. It differs from non-sampling errors, which include measurement error, nonresponse, processing errors, coverage errors and interviewer effects. Increasing sample size generally reduces the random component of sampling variability under an appropriate probability design, but it does not automatically correct systematic bias.
A sample can therefore be very precise about the wrong target if its selection process excludes important parts of the population. Coverage error occurs when the sampling frame does not adequately correspond to the target population. Nonresponse can create bias when those who do not participate differ systematically from respondents in ways relevant to the study.
Sampling frames are operational lists or structures from which samples are drawn. A frame might consist of a population register, address database, business directory, geographical grid, administrative record or list of eligible experimental units. Frame quality matters because omitted units cannot be selected through that frame, while duplicate entries can create unequal selection opportunities. In practice, the target population, accessible population and sampled population may differ, and these distinctions should be documented. Researchers may use stratification, oversampling or unequal selection probabilities to ensure adequate observations for small but important subgroups. Survey weights can then compensate for unequal selection probabilities and certain forms of nonresponse, although weighting does not automatically remove all bias.
Sample size determination depends on the objective, design, variability of the outcome, desired precision, confidence level, expected response rate and statistical model. For estimating a proportion under simple random sampling, sample-size calculations can use a specified or conservative anticipated proportion and desired margin of error. More complex designs require design effects and adjustments for clustering, stratification and weighting.
Power calculations are used for many hypothesis-testing and experimental-design problems and depend on the effect size considered important, variability, significance threshold and desired probability of detecting the effect. A sample size selected without reference to the intended analysis can be inefficient or inadequate. Conversely, collecting far more observations than necessary can impose unnecessary cost without resolving systematic sources of bias.
Sampling is used extensively in surveys, polling, quality control, environmental monitoring, medical research, auditing and scientific fieldwork. In industrial quality control, destructive testing makes complete inspection impossible, so samples provide evidence about production processes. In environmental science, measurements may be taken from selected sites or times because continuous observation of every location is impractical. In household surveys, multistage and stratified designs can make national coverage feasible while controlling costs. In laboratory research, biological or technical replicates provide information about variability, but repeated measurements of the same experimental unit are not automatically independent replicates. Correctly identifying the sampling unit and the experimental unit is therefore essential for appropriate analysis.
Good sampling practice begins by specifying the research question, target population, sampling unit, eligibility criteria, sampling frame and selection procedure before observations are collected. The design should anticipate nonresponse, exclusions, missing observations and possible changes in population structure. Documentation should record how units were identified, selected, contacted and replaced if replacement was permitted. Researchers should distinguish random sampling from random assignment and should not describe a convenience sample as representative merely because it is large. Transparent reporting allows readers to evaluate external validity and potential selection bias. Sampling is consequently not merely a preliminary administrative step but a fundamental component of research design that determines what population the resulting evidence can legitimately describe.
