Probability Sampling

Probability sampling is a sampling method in which every member of the target population has a known, nonzero probability of being selected for the study. This property enables statistical inference from the sample to the population, allowing researchers to calculate margins of error and confidence intervals. The major types of probability sampling include simple random sampling, systematic sampling, stratified sampling, and cluster sampling.

The ability to generalize findings from a sample to a broader population depends on whether the sample was drawn using a method that gives each population member a calculable chance of selection. Probability sampling is the foundation of survey research, epidemiological studies, and any research design that aims to produce population level estimates. Without it, the sample may not represent the population, and any statistics calculated from the sample may be biased in unknown ways and directions. National health surveys and large government surveys such as the American Community Survey rely on probability sampling methods to produce valid inferences; many election polls now draw respondents from online opt-in panels instead, so their estimates rest on weighting and modeling assumptions rather than on the sampling design.

Simple random sampling gives every possible sample of a given size the same probability of being drawn, so every individual has an equal probability of selection; it is typically done using a random number generator applied to a complete list (sampling frame) of the population. While conceptually pure, it requires a complete enumeration of the population, which is often impractical.

Systematic sampling selects every kth individual from an ordered list, starting from a randomly chosen point. It is simpler to implement than simple random sampling and produces similar results unless there is a periodic pattern in the list that coincides with the sampling interval.

Stratified sampling divides the population into mutually exclusive subgroups (strata) based on a characteristic relevant to the research question (such as age group, sex, or geographic region) and then draws a random sample independently from each stratum. This ensures representation of each subgroup and can increase precision compared to simple random sampling, particularly when strata are homogeneous internally but heterogeneous across groups.

Cluster sampling selects entire groups (clusters) of individuals rather than individuals themselves. Researchers first sample clusters (such as schools, clinics, or geographic areas) and then sample individuals within selected clusters, or include all individuals in the selected clusters. Cluster sampling is efficient when no complete list of individuals exists but lists of clusters are available. However, because individuals within clusters tend to be more similar than individuals across clusters, the effective sample size is smaller than the raw count, quantified by the design effect. For clusters of equal size m, the design effect is approximately 1 plus (m minus 1) times the intracluster correlation coefficient (ICC): with 40 people per cluster and an ICC of 0.02, it is 1 + 39 × 0.02 = 1.78, so 1,200 people give about as much information as 674 sampled by simple random sampling.

Each method has tradeoffs between feasibility, cost, and statistical precision. In practice, large surveys often use multistage designs that combine several probability sampling methods, for example sampling counties, then census blocks within them, then households, then one adult per household. A person's overall probability of selection is the product of the probabilities at each stage, and the base survey weight is its inverse: if a county is selected with probability 0.2, a block within it with probability 0.1, a household within the block with probability 0.05, and one of the household's two adults is chosen (0.5), that adult's probability is 0.2 × 0.1 × 0.05 × 0.5 = 0.0005, and the base weight is 2,000, so the respondent stands for about 2,000 adults. Probabilities can differ between people, as long as each is known and above zero.

Probability sampling does not guarantee that any single sample mirrors the population: by chance, a particular sample can be unrepresentative. What it provides is the basis for design based inference: because every selection probability is known, weighted estimates are unbiased on average across repeated samples drawn the same way, and their sampling error can be calculated from the design.

Use probability sampling whenever the research goal is to make statistical inferences about a defined population. It is essential for prevalence surveys, survey research, and any study that requires representative estimates. When a complete sampling frame does not exist, cluster or multistage sampling provides a practical probability based alternative. If the goal is exploratory, hypothesis generating, or focused on rare populations that cannot be enumerated, nonprobability methods such as convenience or snowball sampling may be necessary, though the resulting estimates cannot be generalized with known precision.

A major error is treating a convenience sample as if it were a probability sample, applying inferential statistics and confidence intervals to data that do not support population level generalization. Another mistake is constructing an incomplete or outdated sampling frame, leading to undercoverage of certain population segments. In cluster sampling, researchers often forget to account for the design effect in their power calculations and analyses, producing overly optimistic precision estimates. Finally, non response bias, which occurs when selected individuals do not participate, can undermine even a well designed probability sample if response rates are low and nonrespondents differ systematically from respondents.

A national health survey aims to estimate the prevalence of hypertension among adults aged 18 and older. The sampling design uses a multistage probability approach: first, 120 counties are selected with probability proportional to population size; second, census blocks within selected counties are sampled; third, households within blocks are randomly selected; and finally one adult per household is randomly selected and interviewed. The design achieves a 72% response rate across 28,000 completed interviews. Survey weights adjust for unequal selection probabilities and nonresponse, and the 95% confidence intervals are calculated from the design (its strata and clusters), not as if the 28,000 adults were a simple random sample. With 28% of selected adults not responding, the estimates are nationally representative only to the extent that the weights capture how nonrespondents differ from respondents.