Simple Random Sampling
Simple random sampling (SRS) is a sampling design in which a fixed number n of units is drawn from a population of N units so that every possible subset of size n has exactly the same chance of being the sample. That requirement is stronger than saying every individual has an equal chance of being picked: it also means every combination is equally likely, which is what makes the design's statistical behavior fully calculable. With N = 5 people and n = 2 there are 10 possible pairs, and in an SRS each pair has probability 1/10. In practice SRS is usually done without replacement, so no unit appears twice, and it is the baseline against which every more elaborate design, such as stratified, cluster or systematic sampling, is compared.
SRS is the design for which the textbook formulas hold exactly, and it is what lets a sample speak for a population without any model of how the population is arranged. Under SRS the sample mean is an unbiased estimator of the population mean, the sample proportion is unbiased for the population proportion, and the variability of those estimates across possible samples (their standard-error) can be computed from the sample itself. Because it relies only on the physical act of random selection, not on assumptions about the data, SRS gives design-based inference: the probabilities come from the sampling procedure, not from a model. This distinguishes it from convenience or volunteer samples, where no probability statement about representativeness is justified however large the sample, and it relates to the general idea of random-sampling by pinning down one specific, well-defined version of it.
Executing an SRS requires a sampling frame, a list or mechanism that identifies every unit of the population, and a genuine randomizing device. Historically that meant printed tables of random digits (such as the RAND Corporation's 1955 volume A Million Random Digits); today it means a random number generator that shuffles the frame and takes the first n records, or draws n distinct row numbers. Frame quality matters as much as the draw itself: a perfectly random selection from an incomplete or skewed list yields a sample of the list, not of the population. Typical applications include auditing a random subset of transactions, drawing records for quality inspection from a production lot, selecting survey respondents from a registry, and creating holdout sets in machine learning, where a random split is the SRS idea applied to rows.
A sample's size cannot rescue a sample's design, and the best-known demonstration is the Literary Digest poll of 1936. The magazine mailed ballots to millions of people drawn largely from telephone directories and automobile registration lists, received roughly 2.4 million replies, and predicted that Alf Landon would defeat Franklin Roosevelt; Roosevelt won in a landslide. In the Depression, those lists over-represented wealthier households and the respondents were self-selected, so the huge sample was precisely wrong, while George Gallup's much smaller, more carefully designed sample came close to the actual result. The lesson is that a random mechanism over a complete frame, not volume, is what protects against bias, and that nonresponse can undermine even a well-drawn SRS.
Two further points keep SRS in perspective. First, random does not mean haphazard: standing at a street corner and picking whoever seems convenient is not random, and the human tendency to choose representatively is itself a bias. Second, randomness guarantees representativeness only in expectation: any single SRS can by chance over-represent some group, which is why uncertainty is reported, and why when sampling is a sizeable fraction of a finite population (say more than about 5 to 10%) the standard error is multiplied by the finite population correction, the square root of (N − n)/(N − 1). When subgroups matter, designers often move to stratified sampling to guarantee their representation instead of leaving it to chance.
Read simple random sampling as the clean, equal-chance draw that makes sample-to-population inference a matter of probability rather than faith: every possible sample is equally likely, so the estimates are unbiased and their uncertainty calculable, provided the frame is complete and the selected units actually respond.