Most studies cannot include everyone, so researchers study a sample and draw conclusions about a larger population. The sampling method decides whether those conclusions are trustworthy. A good method gives every part of the population a fair chance of being included; a poor method quietly studies only the easiest people or animals to reach, and the results describe them, not the population.
This guide explains how to choose a sampling method for a research study. It covers probability and non-probability approaches, matching the method to the design, sample size basics, sampling in animal and field studies, common mistakes, and honest reporting. It builds on our guides to writing a methodology section and choosing the right statistical test, and it pairs with our guide to writing results and discussion sections.
A sample is useful only if it resembles the population it stands for. When the sample differs systematically from the population, the study suffers from selection bias: the results reflect who was included, not what is true. A survey of dairy farmers conducted only at a veterinary hospital, for example, will over-represent farmers who already seek veterinary care, and its findings cannot be generalised to all farmers.
Three terms keep the discussion clear. The target population is the whole group the study wants to describe, such as all smallholder dairy farmers in a province. The sampling frame is the list or source from which the sample is actually drawn, such as a registry of farms. The sample is the group finally studied. Problems arise whenever the frame misses part of the population, because those people or animals can never be selected, however good the method.
Sampling also affects cost and feasibility. Studying everyone is usually impossible, and studying a well-chosen sample is both cheaper and faster than a poorly planned attempt to study many. The goal is not the largest sample, but a sample chosen by a method that matches the research question and can be defended to reviewers.
In probability sampling, every member of the population has a known, non-zero chance of being selected, and selection involves randomness. This is the strongest basis for generalising results, because random selection tends to balance out unknown differences between the sample and the population. The main methods differ in how the randomness is organised.
Simple random sampling gives every member an equal chance, usually by drawing names or numbers at random from the frame. It is the cleanest method, but it needs a complete frame, which is not always available. Systematic sampling selects every nth member from an ordered list, which is simpler to carry out, but it can go wrong if the list has a hidden pattern that lines up with the interval.
Stratified sampling divides the population into subgroups, called strata, such as farm size or region, and samples randomly within each one. It guarantees that important subgroups are represented, even small ones, and it usually gives more precise estimates. Cluster sampling goes the other way: it samples whole groups, such as villages or herds, and studies everyone within the chosen groups. It is practical when the population is spread out, but members of one cluster tend to resemble each other, so the effective sample is smaller than the headcount suggests.
Multistage sampling combines these ideas for large, spread-out populations: first sample districts, then villages within them, then households or herds within those villages. Each stage uses random selection, which keeps the design a probability design, but the analysis must account for the stages, since precision falls at each level of clustering.
In non-probability sampling, selection is not random, and the chance of inclusion is unknown. These methods are common in exploratory work, qualitative studies, and situations where no frame exists. Their weakness is that results cannot be generalised statistically; their strength is that they make studies possible where random sampling cannot.
Convenience sampling takes whoever is easiest to reach: patients at one clinic, students in one class, or farmers at one market. It is fast and cheap, but it is the method most open to bias, and its limits must be stated plainly. Purposive sampling selects participants deliberately because they have particular knowledge or experience, such as experienced veterinarians for a study of clinical judgement. It suits qualitative research, where depth matters more than representativeness.
Quota sampling fills set numbers from each subgroup without random selection within them, while snowball sampling asks participants to refer others, which is useful for hidden or hard-to-reach groups. Both are practical, but both carry bias: quota samples reflect the recruiter's choices, and snowball samples reflect social networks. Neither supports the statistical generalisations that probability sampling does.
The research question decides the method. A prevalence survey that must estimate a rate for a whole region needs probability sampling, because only random selection supports that kind of generalisation. An exploratory interview study that seeks to understand experiences in depth needs purposive sampling, because random selection would waste interviews on people with nothing relevant to say.
Experimental studies need a different distinction: random sampling selects who enters the study, while random assignment divides participants into treatment and control groups. Both matter, but they solve different problems. Random assignment balances the groups being compared; random sampling supports generalising beyond the study. A trial can have one without the other, and the methodology section should describe each separately.
Feasibility also belongs in the decision. A perfect method that cannot be carried out with the available time, budget, or access is not a good method for that study. The honest approach is to choose the strongest method that is actually feasible, and to report its limitations clearly rather than pretending the sampling was random when it was not.
Document the decision and its reasons in the study protocol before sampling begins. A written rationale protects the study when reviewers ask why a particular method was chosen, and it prevents mid-study changes driven by convenience rather than design. If the method must change during fieldwork, record the change and its reason immediately.
Sample size should be decided before data collection, not after seeing the results. For studies that estimate a proportion or compare groups, a formal calculation uses the expected effect or prevalence, the variability, the desired confidence level, and the power. Many textbooks and free calculators cover these formulas, and the methodology section should report the inputs, not only the final number.
For qualitative studies, sample size follows a different logic: data are collected until new interviews stop adding new themes, a point called saturation. Claiming saturation after three interviews is not credible, but neither is a fixed number copied from another study without justification. The study should explain how sufficiency was judged.
Practical adjustments matter too. Allow for non-response and dropouts by inflating the target sample, and remember that cluster sampling needs a larger headcount than simple random sampling for the same precision. A sample size that looks adequate on paper but ignores these realities will end up too small in practice, and an underpowered study wastes everyone's effort.
Animal and field studies add complications that human surveys do not. Animals are grouped in herds, flocks, and pens, so the herd is often the natural cluster, and sampling animals within a few herds is not the same as sampling animals across many herds. Disease also clusters: animals in one herd share exposures, so treating clustered animals as independent observations overstates the real sample size.
Field conditions also limit the frame. A registry of farms may be incomplete, seasonal migration may move animals, and some farms may refuse access. Researchers should describe how the frame was built, which areas or herds were excluded and why, and how refusals were handled. These details let reviewers judge whether the sample can support the conclusions.
Welfare and ethics constrain sampling as well. Repeated handling, transport, or invasive sampling must be justified, and the number of animals sampled should be no larger than the study needs. Our guide to research ethics and animal welfare explains the approval and welfare requirements that apply before any animal sampling begins.
The most common mistake is confusing the sample with the population in the conclusions. A study of farmers in one district cannot support claims about farmers nationwide, and a study of one breed cannot support claims about all cattle. Conclusions must stay within the population the sample can actually represent.
Another frequent mistake is silent substitution: replacing participants or animals that drop out with convenient replacements, without reporting it. Substitution changes the sample's composition and can reintroduce exactly the bias the original method was designed to avoid. If substitution happens, it must be reported and its effect discussed.
Finally, many papers describe their sampling vaguely, with phrases like "randomly selected" that do not explain how. Vague descriptions prevent reviewers from judging the method, and they invite suspicion. A precise description, even of a simple method, is far more convincing than a grand claim without detail.
A complete sampling description answers six questions: what was the target population, what was the frame, which method was used, how was it carried out step by step, how many were invited, selected, and finally studied, and what were the response and exclusion rates. A flow diagram of these numbers, similar to those used in clinical trials, makes the account easy to follow.
Limitations belong in the same account. If the frame missed part of the population, if the response rate was low, or if clustering reduced the effective sample, say so and discuss what it means for the conclusions. Reviewers accept honest limits; they do not accept a sampling section that hides them, as our guide to writing the limitations section explains.
Keep the language precise throughout. "Random" should mean a random process was used, not merely that selection felt unsystematic. "Representative" should mean the sample resembles the population on the characteristics that matter, with evidence, not that the author hopes it does. Precise language in the methods section is what makes the results section believable.
Finally, archive the sampling records with the study data: the frame, the randomisation output, the invitation log, and the exclusion list. Good records let other researchers verify the sampling, support secondary analysis, and answer reviewer queries long after fieldwork ends, which is exactly what our guide to research data management recommends.
| Method | Type | How it works | Main strength and limit |
|---|---|---|---|
| Simple random | Probability | Equal random chance for all | Cleanest method; needs a full frame |
| Systematic | Probability | Every nth member from a list | Simple; hidden list patterns can bias it |
| Stratified | Probability | Random sampling within subgroups | Guarantees subgroup representation; more planning |
| Cluster | Probability | Random whole groups, all members studied | Practical for spread populations; less precision |
| Convenience | Non-probability | Whoever is easiest to reach | Fast and cheap; most open to bias |
| Purposive | Non-probability | Deliberate choice for knowledge or experience | Depth for qualitative work; no statistical generalisation |
| Snowball | Non-probability | Participants refer further participants | Reaches hidden groups; reflects social networks |
These authoritative external resources explain the standards and guidance mentioned in this article:
What is the difference between probability and non-probability sampling?
In probability sampling, every member has a known, non-zero chance of selection through randomness, which supports statistical generalisation. In non-probability sampling, selection is not random, so results describe the sample, not necessarily the whole population.
How do I decide my sample size?
For quantitative studies, calculate it before data collection from the expected effect or prevalence, variability, confidence level, and power, and report the inputs. For qualitative studies, collect data until new material stops adding themes, and explain how that point was judged.
Is convenience sampling acceptable?
It is acceptable for exploratory or pilot work and where no sampling frame exists, but its limits must be stated plainly. It cannot support broad generalisations, and conclusions must stay within what the sample can represent.
What is the difference between random sampling and random assignment?
Random sampling selects who enters the study from the population; random assignment divides participants into treatment and control groups. Sampling supports generalisation, while assignment balances the groups being compared. A study can have one without the other.
Why does clustering matter in animal studies?
Animals in one herd or flock share exposures, so they are not independent observations. Treating clustered animals as independent overstates the real sample size and can produce misleadingly precise results. The analysis should account for clustering.
What should the methodology section say about sampling?
It should state the target population, the sampling frame, the method, the step-by-step procedure, the numbers invited, selected, and studied, and the response and exclusion rates, plus the limitations of the approach chosen.