A company wants to know how many customers encountered a billing problem last month. Asking whoever replies to a newsletter may produce useful accounts, but the selection process is difficult to describe: the team does not know each customer’s chance of appearing in the result. Probability sampling begins by making that selection mechanism explicit.
In a probability sample, units are selected through a random mechanism with known, non-zero inclusion probabilities for the population covered by the design. The probabilities need not be equal. Their role is to support estimation and an assessment of sampling uncertainty that reflects how the sample was drawn.
Define the population and the route into the sample
In the fictional billing study, the population might be all accounts billed during a specified month. The customer database can provide a sampling frame if it adequately lists those accounts, but it may not identify every individual who experienced the billing process. Choose the unit that matches the question before drawing the sample.
Simple random sampling selects directly from a frame. Stratified sampling draws within defined groups, while cluster and multistage designs select through groups such as organisations or locations. The Australian Bureau of Statistics’ sampling overview explains the distinction between probability and other sampling approaches.
Document every stage that affects inclusion. Oversampling a small customer group can be useful for analysis, but an overall estimate must account appropriately for the unequal probabilities rather than treating the sample composition as the population composition.
Keep selection separate from response
A random invitation does not guarantee that everyone selected will answer. The achieved responses can differ from the sampled group, so follow-up, weighting and analysis of missing responses may be necessary. The study remains based on a probability design, but its estimates depend on how nonresponse is handled and on the assumptions that remain.
A conventional margin of error describes sampling uncertainty under the relevant design and calculation. It does not include every possible error from missing coverage, misleading questions or nonresponse. Reporting a narrow interval therefore does not establish that the survey is accurate overall.
Use the design that fits the claim
Probability sampling is particularly valuable when the study seeks population estimates and a suitable frame is available. It is not required for every useful research question. Purposive sampling can be appropriate for investigating particular experiences or contexts in depth.
Non-probability samples may also support inference through modelling or other adjustments, but that requires additional assumptions and evidence. Matching a few demographic quotas does not recreate a random design or justify using its uncertainty calculation without qualification.
The report should explain the population, frame, selection, response and analysis. That chain lets a reader assess what the estimate describes, rather than relying on “random” as a general assurance of quality.