The population is the full set to which inferences are desired; the sample is the subset studied. Probability sampling (random, stratified, systematic) provides known inclusion probabilities and supports statistical inference. Non-probability sampling (convenience, purposive) sacrifices representativeness but is practical. Sample representativeness on key variables enhances generalizability; bias in sampling limits it.
Specify a research population and design probability and non-probability sampling schemes. Evaluate how different sampling methods affect validity and generalizability. Calculate what percentage of a population you've sampled.
Every empirical study begins with a fundamental question: who are we studying, and who do we want our conclusions to apply to? The population is the full set of individuals, events, or observations to which you want to generalize. The sample is the subset you actually study. The relationship between them — and how that relationship was constructed — determines the generalizability of everything you conclude.
Probability sampling methods give each member of the population a known, nonzero probability of being selected. Simple random sampling is the foundation: every individual has an equal chance of selection, which allows you to use standard inferential statistics — sampling distributions, confidence intervals, and significance tests — in their textbook form, because the theoretical assumptions underlying those tools require random selection. Stratified sampling divides the population into subgroups (strata such as age, gender, or region) and samples randomly within each, ensuring proportional representation of key variables even with moderate sample sizes. Systematic sampling selects every kth person from an ordered list — practical when a list exists and the ordering is unrelated to the outcome. All probability methods share a crucial property: selection is governed by chance, not judgment, so sample estimates are unbiased estimates of population parameters.
Non-probability sampling includes convenience samples (whomever is available and willing), purposive samples (selected for specific characteristics), and snowball samples (participants recruit others from their networks). These are common in psychology — undergraduate participant pools, clinic populations, online panels — and they are not inherently invalid. But they require honesty about generalizability limits. A convenience sample of university students may generalize well to similar populations but poorly to older adults, rural communities, non-Western cultures, or people with limited education. The now-famous WEIRD critique (Western, Educated, Industrialized, Rich, Democratic) of psychology's participant base is fundamentally a sampling argument: if most published research draws from a narrow slice of humanity, conclusions cannot be claimed to describe universal human psychology.
The most important misconception to overcome is confusing large sample size with representativeness. A famous historical example: the Literary Digest predicted a landslide win for Alf Landon in the 1936 US presidential election based on a survey of 10 million people — one of the largest polls ever conducted. Franklin Roosevelt won by the largest Electoral College margin in over a century. The poll failed not because it was too small but because it drew from telephone directories and car registration lists, systematically oversampling wealthy, Republican-leaning voters. Representativeness is a question of *who* is selected and how, not how many. A well-drawn random sample of 1,000 routinely outperforms a convenience sample of 1,000,000 for accurate population inference.