Populations versus Samples

A population (also called the target population or statistical population) is the entire, precisely defined collection of individuals, items, or events about which a researcher wants to draw a conclusion. It need not be a biological population in the ecological sense of "all coyotes in Yellowstone" — in statistics, "population" is an abstract term that can refer to all the leaves on one oak tree, all the possible offspring from a genetic cross, every heartbeat a patient will ever have, or every possible replicate of an experiment that could in principle be run. What makes a population a population is that it is the complete set of units the conclusion is meant to apply to — nothing more restrictive, nothing more expansive. A sample is the subset of that population actually observed or measured. The sample is not interesting in itself; it is a means to an end. Biologists rarely care what the 200 captured finches' beaks measured — they care what that tells them about all Daphne Major finches, or all medium ground finches generally. This is the core epistemic move of inferential statistics: use measurable, finite sample data to make justified claims about an unmeasurable, and often infinite or hypothetical, population. A crucial and frequently misunderstood point is that the population is defined by the scientific question, not by the biology alone. Consider a single field of 500 wheat plants. If the question is "what is the mean height of wheat in this field this year," the population is those 500 plants, and a full census is possible. But if the question is "what is the mean height wheat plants of this genotype reach under this fertilizer regime," the population becomes every wheat plant of that genotype that could ever be grown under those conditions — an infinite, hypothetical population — and the 500 actual plants are now merely a sample from it. The same physical plants can be a population under one framing and a sample under another; the statistical treatment (and the honesty of the conclusion) depends entirely on which framing the scientist has chosen and stated. This links tightly to Types of Biological Data, since the kind of data collected (counts, measurements, categories) constrains what population-level summaries are even meaningful.

Step 1: Define the target population precisely, before collecting any data. This means specifying the biological entity (species, tissue, allele, patient group), the spatial or temporal extent (this lake in July 2024; adults over 18 with type 2 diabetes), and any inclusion/exclusion criteria. A vague population definition ("fish in the ocean") cannot be sampled meaningfully because it is not falsifiable or bounded; a precise one ("adult yellow perch, Perca flavescens, in Lake X between May and September 2024") can be.

Step 2: Identify the sampling frame. The sampling frame is the actual, practical list or mechanism through which units can be accessed — nets set in the lake, patients arriving at a clinic, tissue samples in a freezer archive. The sampling frame is almost never identical to the target population; a lake's fish that swim near the shore where nets are set are not the same as all fish in the lake, and hospital patients are not the same as all people with a disease (many never seek care). Any systematic gap between sampling frame and target population is a source of sampling bias, distinct from ordinary sampling error.

Step 3: Draw the sample using a probability-based design. Simple random sampling gives each unit in the frame an equal, independent chance of selection. Stratified sampling divides the population into subgroups (e.g., by age class, habitat type, disease stage) and samples within each to guarantee representation. Systematic and cluster sampling are practical compromises used when true randomization is logistically impossible (e.g., sampling every tenth tree along a transect, or sampling whole nests of birds rather than individual chicks). The mathematical guarantees of later inferential statistics — unbiased estimators, correct standard errors, valid confidence intervals — depend on this randomization step having actually occurred.

Step 4: Measure the sample and compute sample statistics. Data are collected on the sampled units according to the data types established in Types of Biological Data (continuous, discrete, categorical, ratio, ordinal). Summary numbers computed from the sample — a sample mean, sample proportion, sample variance — are called statistics, and they are used as estimates of the corresponding parameters, the true (usually unknown and unknowable) summary values of the full population. This estimate/target relationship is the subject that Parameters versus Statistics develops fully; here it is enough to see that the sample only has value insofar as it stands in for the population.

Step 5: Generalize back to the population, with explicit uncertainty. Because the sample is not the whole population, any number calculated from it differs somewhat from the true population value purely by chance — this gap is sampling error, and it shrinks (on average) as sample size grows, following patterns rigorously described later. Sampling error is unavoidable and is not a mistake; it is the mathematically expected consequence of measuring a part rather than the whole. Sampling bias, by contrast, is a systematic distortion caused by a flawed sampling frame or non-random selection, and it does not shrink with a larger sample — it can only be fixed by redesigning how the sample is drawn.

The Literary Digest poll of 1936 is the canonical demonstration, from outside biology but foundational to how biologists now think about sampling, of why sample size alone guarantees nothing. The magazine mailed 10 million questionnaires to names drawn from telephone directories and automobile registration lists, and received about 2.4 million responses — a colossal sample by any standard. It predicted Alf Landon would defeat Franklin D. Roosevelt for the U.S. presidency. Roosevelt won in a landslide, carrying 46 of 48 states. The sampling frame — phone and car owners in 1936, during the Great Depression — systematically over-represented wealthier voters, who leaned Republican, while excluding poorer voters, who leaned Democratic and voted for Roosevelt. Simultaneously, George Gallup, using a far smaller but carefully constructed sample of a few thousand respondents chosen to represent the demographic structure of the electorate, correctly predicted the Roosevelt win. The lesson generalizes directly to biology: a large sample built on a bad sampling frame reproduces its bias at scale, while a modest, well-designed random or stratified sample tracks the population accurately. This episode is why modern experimental design courses insist that biologists specify the sampling frame and selection method before worrying about sample size.

Jerzy Neyman's 1934 paper "On the Two Different Aspects of the Representative Method" provided the formal mathematical foundation that biology's sampling practice now rests on. Before Neyman, "representative sampling" often meant a researcher subjectively chose units that seemed typical — a method with no guarantee of accuracy and no way to compute how much error to expect. Neyman proved that only probability sampling — where each unit's chance of inclusion is known — allows valid computation of confidence intervals and standard errors, and he formalized stratified random sampling, showing it can be more efficient (lower variance for the same sample size) than simple random sampling when the population contains identifiable subgroups with different characteristics. This work, alongside R. A. Fisher's contemporaneous development of randomization in agricultural field trials, established the mathematical machinery that every population-versus-sample study in modern ecology, clinical trials, and genetics now uses, whether explicitly stated or not.

A biological illustration: the Christmas Bird Count and population monitoring. Ornithologists cannot count every individual bird of a species across a continent. The National Audubon Society's Christmas Bird Count, running continuously since 1900, instead defines fixed 24-km-diameter survey circles, sends volunteers out on a designated day each year, and uses consistent counting protocols within those circles as the sampling frame. The resulting counts are not the true continental population size, but because the same circles and broadly similar effort are used year after year, changes in the sample track changes in the population reliably enough to have documented, for example, sharp multi-decade declines in many grassland bird species. This shows sampling used correctly: the goal was never to census all birds, but to build a consistent, repeatable sample structure from which trends in the true population can be inferred with quantifiable confidence — directly foreshadowing the logic behind confidence intervals and hypothesis tests built on sample statistics.

Example 1 — Ecology: estimating a wolf population from track surveys. A wildlife biologist wants to know the total number of wolves in a 3,000 km² management unit — this unbounded, unknowable number is the target population parameter, N. She cannot count every wolf directly (wolves are elusive, nocturnal, and range widely), so she uses a mark-recapture design as her sampling method.

• First survey (marking): 28 wolves are captured, fitted with GPS collars, and released. (n₁ = 28)
• Second survey, several weeks later: 40 wolves are observed via camera traps and collar signals. (n₂ = 40)
• Of those 40, 7 carry collars from the first survey. (m = 7)

Using the Lincoln–Petersen estimator, a population size estimate:

This 160 is a sample-derived estimate (a statistic) of the true, unknown population size N (a parameter). It is not exact — a slightly different second survey would yield a slightly different estimate purely from sampling error, and the biologist would normally compute a confidence interval around 160 to express that uncertainty. Note also the sampling frame problem: if wolves near roads or open terrain are more easily photographed than wolves deep in dense forest, the sample is biased toward "detectable" wolves, and 160 could systematically underestimate or overestimate the true population depending on which habitat holds the untracked animals — exactly the Literary Digest problem, transplanted into a forest.

Example 2 — Clinical genetics: allele frequency estimation and study population definition. A geneticist studying cystic fibrosis wants to estimate the frequency of the ΔF508 allele of the CFTR gene in a European-ancestry population. The target population is, in principle, "all people of European ancestry" — billions of people, impossible to genotype in full. The actual study samples 500 unrelated individuals recruited from clinics in three countries. Genotyping finds:

• 480 individuals with genotype AA (no ΔF508 allele)
• 18 individuals with genotype Aa (one copy)
• 2 individuals with genotype aa (two copies, cystic fibrosis-affected)

Total alleles sampled = 500 × 2 = 1,000. Count of ΔF508 (a) alleles = (18 × 1) + (2 × 2) = 22.

This 0.022 is a sample statistic estimating the true population allele frequency q. Two distinct issues must be flagged before generalizing this number to "Europeans" broadly: (1) sampling error — a different sample of 500 would give a somewhat different estimate, quantifiable with a standard error and confidence interval (developed under Parameters versus Statistics); and (2) sampling bias — because the participants were recruited from clinics, and cystic fibrosis families are disproportionately likely to seek clinical genetic testing (ascertainment bias), the sampling frame may over-represent carrier families relative to the true general population, inflating the estimate of q above its true value in unaffected relatives of patients. A geneticist reporting this number responsibly states the actual recruited population ("clinic-referred European-ancestry individuals in three countries, 2020–2023") rather than silently generalizing to "Europeans" at large.

Jerzy Neyman (1894–1981) was a Polish-American mathematician who transformed sampling from an intuitive craft into a rigorous mathematical discipline. Trained in Poland and Russia and shaped by the statistical seminars of Karl Pearson's laboratory in London, Neyman arrived at University College London in the early 1930s at a moment when official statistics agencies across Europe relied heavily on "purposive selection" — experts hand-picking districts, farms, or households they judged to be typical of the whole. This method had no mathematical guarantee of accuracy and offered no way to compute how far a sample estimate might lie from the truth. Neyman's 1934 paper to the Royal Statistical Society demonstrated, with both mathematical proof and worked agricultural examples, that only probability-based sampling — where selection follows known chance mechanisms rather than expert judgment — permits valid calculation of standard errors and confidence intervals for population estimates. He introduced and formalized stratified sampling, proving that dividing a heterogeneous population into more homogeneous strata before sampling within each stratum can substantially reduce the variance of an estimate for a given total sample size — a technique now standard in ecological surveys (stratifying by habitat type), clinical trials (stratifying by disease severity), and census design worldwide. Neyman went on to co-develop, with Egon Pearson, the Neyman–Pearson framework for hypothesis testing that underlies the logic of statistical significance used throughout biology, and later moved to the University of California, Berkeley, where he built one of the twentieth century's most influential statistics departments and mentored generations of applied statisticians working on biological and medical problems.

Clinical trials and drug approval. Regulatory agencies require trial sponsors to pre-specify the target population (e.g., "adults aged 18–75 with HbA1c between 7% and 10%") before enrollment, precisely so that later generalization of results is honest. The recurring finding that many landmark cardiovascular and cancer trials historically under-enrolled women, older adults, and non-white participants is a sampling-frame problem: the drug's effect in the study sample may not generalize cleanly to the full population of patients who will actually receive the drug in practice — a live concern shaping current FDA and EMA diversity-in-trials guidance.

Conservation and wildlife management. Endangered species listings, hunting quotas, and habitat protection decisions depend on population size estimates built from samples — camera-trap grids, transect counts, mark-recapture studies like the wolf example above. Getting the sampling frame wrong (surveying only accessible terrain, or only daylight hours) can systematically bias estimated population trends in either direction, directly affecting whether a species is protected or a harvest quota is raised.

Genetic epidemiology and precision medicine. Genome-wide association studies (GWAS) estimate allele frequencies and disease-risk associations from sampled cohorts, most historically drawn overwhelmingly from populations of European ancestry. Polygenic risk scores built from these samples have been shown to predict disease risk substantially less accurately in individuals of African or Indigenous ancestry, a direct consequence of the sampling frame not matching the target population of "all humans" that clinical tools are ultimately meant to serve — a major active problem in current biomedical research funding priorities.

Agricultural field trials. Testing a new crop variety's yield requires defining the target population precisely — this variety, under what range of soils, climates, and management practices — and then sampling representative field plots rather than a single favorable field, following the randomization principles pioneered by Fisher at Rothamsted, so that yield claims generalize beyond the specific plots tested.

Public health surveillance. Estimates of disease prevalence (e.g., seroprevalence surveys during epidemics) depend on how the sample was recruited; convenience samples of people willing to be tested systematically differ from the general population in ways (symptomatic status, healthcare access, risk behavior) that bias prevalence estimates, a lesson relearned repeatedly during COVID-19 seroprevalence studies worldwide.

Ask the tutor: If my sample was randomly selected but still small, how do I know whether my estimate is close enough to the true population parameter to trust?