A scientific sample is a subset of observations, individuals, objects, measurements or experimental units selected from a broader population or defined study context for the purpose of investigation. Researchers use samples because examining every member of a population is often impossible, impractical, expensive or unnecessary. A population can mean all individuals or units to which the researcher wants to generalise, while a sample is the subset actually observed.
In experimental research, the sample may consist of participants, organisms, materials or physical systems assigned to different conditions. In observational research, it may consist of naturally occurring observations selected through a defined sampling process. The scientific value of a sample depends on how it was selected, how accurately it represents the target population, how variables were measured and whether the sample size is adequate for the intended inference.
Sampling provides the bridge between observed data and conclusions about a larger population. A parameter is a numerical characteristic of the target population, such as its mean or proportion, while a statistic is the corresponding quantity calculated from a sample. Statistical inference uses sample information to estimate population parameters and quantify uncertainty. Because different samples can produce different statistics, sampling variability is unavoidable. A sufficiently well designed random sample can allow researchers to estimate population characteristics with measurable uncertainty.
The standard error and confidence interval are tools for describing aspects of this sampling uncertainty. However, increasing sample size does not automatically correct systematic sampling bias. A very large sample selected through a flawed process can provide a highly precise estimate of the wrong population quantity.
Scientific sampling methods include simple random sampling, systematic sampling, stratified sampling, cluster sampling and multistage sampling. In simple random sampling, units have known selection probabilities and are selected randomly according to the design. Stratified sampling divides a population into relevant strata and samples within them, potentially improving representation of important subgroups. Cluster sampling selects groups or clusters rather than individual units and can reduce logistical costs, although observations within clusters may be correlated.
Multistage designs combine several sampling stages and are common in large population surveys. Convenience sampling selects easily accessible units and is often less defensible for population generalisation because inclusion probabilities may be unknown or systematically related to characteristics of interest. The appropriate design depends on the population structure, research objective, resources and required precision.
A scientific sample can also refer to a laboratory specimen or physical material rather than a statistical subset of a population. In chemistry, a sample might be a portion of material submitted for analysis. In geology, it might be a rock specimen collected from a particular location. In microbiology, it could be a biological specimen obtained from an environment or organism. The term sample therefore has both statistical and laboratory meanings.
In both cases, representativeness can be important, but the relevant concept differs. A laboratory sample must adequately represent the material or process being characterised and must be handled without unacceptable contamination or degradation. A statistical sample must be selected in a way that supports the intended inference. Confusing these meanings can lead to imprecise descriptions of scientific methodology.
Sample size is determined by factors including the expected effect size, variability, desired precision, statistical power, significance level, study design and available resources. Larger samples generally reduce random sampling error and provide more precise estimates, although the improvement depends on the design and information content of the observations. In clustered data, for example, increasing the number of observations within the same cluster may provide less independent information than adding observations from additional clusters because within cluster observations can be correlated.
Power calculations can help determine an appropriate sample size for a planned hypothesis test or estimation problem. A sample that is too small may produce unstable estimates and wide uncertainty intervals, while an unnecessarily large sample can waste resources and may detect extremely small effects that have little practical significance.
Sampling quality also depends on inclusion and exclusion criteria. Researchers define who or what is eligible for the study and how potential units are identified. If eligibility criteria systematically exclude relevant members of the target population, the resulting sample may not support the intended generalisation. Nonresponse can introduce another source of bias when selected units decline participation and nonresponders differ from responders in relevant ways. Attrition in longitudinal research can similarly change the composition of the observed sample over time. Researchers therefore distinguish the target population, sampling frame, recruited sample and analysed sample. These distinctions are important because the population that a study technically represents may be narrower than the broad population implied by an informal description of the research.
Scientific samples require appropriate handling of measurement uncertainty and data quality. Repeated measurements from the same individual or object are not necessarily independent observations, and treating them as independent can artificially increase apparent sample size. Pseudoreplication occurs when multiple measurements are incorrectly treated as though they came from separate independent experimental units. Randomisation of treatment assignment is also distinct from random sampling.
A study can randomly assign participants to treatments while using a convenience sample, which strengthens internal causal inference under appropriate conditions but may limit generalisability. Conversely, a probability sample can represent a population well while providing no basis for causal inference if the study is purely observational. Sampling and experimental design therefore answer different methodological questions.
A scientific sample is ultimately defined by its relationship to the population or system about which the researcher wants to learn. Good sampling begins with a clearly specified target population, a suitable sampling frame or selection process, appropriate inclusion criteria and a sample size consistent with the scientific objective. The sample should then be measured consistently and analysed using methods that reflect its design.
Generalisation requires attention to sampling bias, nonresponse, missing data and differences between the observed sample and target population. A sample is therefore not scientifically useful merely because it is large. Its value depends on how it was selected, what it represents, how independently observations were obtained and whether the resulting data can support the inference being attempted.
