To know the whole pot of soup, you only have to stir it well and taste one spoonful.
Probability, in Stage 33, asked a question about the future: knowing the coin is fair, how likely is tomorrow's flip to land heads? Statistics turns around and faces the past. The data are already in hand — a pile of numbers x₁, x₂, …, xₙ that yesterday recorded — and the job is to read them honestly: to turn that raw pile into a summary, and the summary into a defensible conclusion about a world far bigger than the pile. The very first decision is the one most people skip: which numbers do we even collect? Get that wrong, and no amount of clever arithmetic later can save you.
Every honest claim about the world rests on data. Do teenagers sleep enough? Is this new drug safer? Which way will the vote go? You cannot answer any of these by thinking harder in an armchair; you have to go out and measure. Statistics is the discipline of measuring well, then reasoning from what you measured.
The whole subject runs along one arc, and Stage 34 walks it end to end:
| Step | What you do | Lesson |
|---|---|---|
| DATA | Decide what to know; collect honest numbers by sampling | 34.1 — you are here |
| SUMMARY | Picture the pile (charts, histograms); squeeze it to a center and a spread | 34.2 – 34.4 |
| INFERENCE | Let the sample speak for the population; find relationships; reach a verdict | 34.5 – 34.7 |
This lesson is the foundation of the foundation. Before any mean or median can mean anything, the numbers have to have been gathered fairly. So we begin with the words, and then with the one skill everything else depends on: sampling.
Statistics is the art of reasoning from data: collect → summarize → infer. The honesty of the conclusion can never exceed the honesty of the collection.
Three plain words carry the whole idea. The population is every object you care about — all of them. Each one is an individual. And the sample is the handful you actually examine. Their sizes have names too: the population has size N, the sample has size n, and almost always n is far smaller than N.
The cook's image says it perfectly. To judge a whole pot of soup, you don't drink the pot. You stir it well and taste one spoonful. The pot is the population; the spoonful is the sample; and stirring is what makes the spoonful represent the whole. A statistic measured on the sample — say the spoonful's saltiness, or in symbols the sample mean x̄ — is our best guess at the truth of the whole pot, the population mean μ. (We will make that leap from x̄ to μ precise in 34.5.)
A principal wants the average nightly sleep of the 2,000 students at her school, but has time to ask only 50.
Population: all 2,000 students (N = 2000). Individual: one student. Sample: the 50 she asks (n = 50). The sleep figure she reports — x̄ from those 50 — is an estimate of the true school average μ she never directly measures.
There are exactly two ways to get the numbers. A census measures every individual in the population — you call the whole roll. Sampling measures only a part and lets it speak for the rest. A national headcount every ten years is a census; a phone poll of 1,000 voters is a sample.
Why ever sample, if a census is complete and exact? Because a census is often impossible, ruinously expensive, or even destructive. You cannot test the lifetime of every light bulb a factory makes — you'd have none left to sell. You cannot ask all 50 million voters before an election. Sampling trades a little uncertainty for enormous savings of time, money, and (sometimes) the very objects being tested.
And here is the comforting part, which 34.5 will quantify: a larger fair sample is generally closer to the truth. More spoonfuls, better-stirred, give a steadier taste. But — and this is the hinge of the whole lesson — larger only helps when the sample is fair. A big unfair sample is just a precisely wrong answer.
Census = measure everyone (complete, exact, costly). Sampling = measure a fair part (cheaper, a little uncertain). Bigger fair samples are steadier — fairness first, size second.
What does “fair” actually mean? The gold standard is simple random sampling (SRS): a method in which every individual has an equal chance of being chosen, and every possible group of size n is equally likely. No favorites, no patterns, no hidden tilt.
The classic recipe is the lottery: write all N names on identical slips, crumple them, drop them in a hat, stir, and draw n without looking. For larger lists, statisticians use a random-number table or a computer's random generator: number the population 1…N, then read off random numbers and pick the matching individuals. Either way, the choosing is handed to chance, which has no agenda.
The widget below draws a real simple random sample from a known population of 40 quiz scores whose true mean is exactly μ = 14. Watch two things: the sample mean x̄ lands near 14 but never exactly on it (a sample wobbles), and as you raise n the wobble shrinks — the spoonful tastes more like the pot.
Pick 4 of 12 club members by lottery. Number them 1–12, draw 4 slips: say 3, 7, 8, 11. Every member had the same 412 = 1/3 chance of being on the slip pulled, and every set of four was equally likely. That equal-chance guarantee is the whole point of SRS.
Pure SRS is fair, but it can be unlucky: by chance it might draw almost no one from a small but important group. Two refinements fix this while staying fair.
First split the population into non-overlapping strata — natural layers like men/women, grade levels, or city/countryside — then draw a simple random sample inside each stratum, taking from each in proportion to its size. If a school is 80% day students and 20% boarders, a stratified sample of 20 takes 16 day students and 4 boarders, guaranteeing both layers are heard in the right balance. Use it when the strata genuinely differ on what you're measuring.
Order the population (a list, a production line), pick a random start, then take every k-th one — every 10th name on the roll, say. It's quick and spreads the sample evenly across the list. The one danger: if the list has a hidden rhythm that lines up with k (every 10th house is a corner lot), the “system” can sneak in a bias.
The toggle below compares a careful stratified allocation with a single careless SRS on two unequal strata. Stratified always honors the 80/20 split; one careless SRS can wildly under- or over-represent the small stratum.
Stratified — split into layers, draw each in proportion (use it when layers differ). Systematic — every k-th after a random start (quick, but beware hidden rhythms). Both build on the equal-chance fairness of SRS.
A sample is good only when it is representative — when its makeup mirrors the population's. When the method systematically tilts the sample one way, we call it bias, and the result is a sample that lies in a predictable direction.
Picture surveying student heights by standing at the door of the basketball gym. Your spoonful is drawn only from the tallest part of the pot, so your average comes out far too high. The number you report is precise, careful, decimal-pointed — and wrong. Two traps cause most real-world bias:
The widget below makes bias visible. A fixed population of 30 students has a known true mean height μ = 165 cm. A fair frame samples the whole school; the basketball-team frame and the volunteers frame are skewed — and the bias (sample mean − μ) glows red.
This is the misconception that bites hardest. Doubling a biased sample doesn't pull it toward the truth — it just makes a precise wrong answer. Measuring 1,000 people at the basketball gym still reports a height far above μ; you've only nailed down the wrong number more tightly. The cure for bias is a fair method, not a bigger one. (Try the “make it bigger” button above — the bias does not budge at all.)
| Term / method | What it means | Why it matters |
|---|---|---|
| Population (N) | every individual you care about | the truth you want, mean μ |
| Sample (n) | the part you actually measure | gives the estimate x̄ of μ |
| Census vs sampling | measure all vs measure a part | sample when census is too costly |
| Simple random | every individual an equal chance | the gold standard of fairness |
| Stratified | layers sampled in proportion | guarantees every layer is heard |
| Systematic | every k-th after a random start | fast; beware hidden rhythms |
| Bias | a method that tilts the sample | a bigger sample never fixes it |
One sentence to remember: fairness first, size second. A fair handful, well stirred, speaks for the whole pot; a tilted handful only speaks for itself, no matter how large.
Population = all 1,250 backpacks, so N = 1250. Individual = one backpack. Sample = the 60 weighed, so n = 60. The average of those 60 is an estimate of the true population mean μ.
A sample. Measuring lifetime is destructive — a battery tested to failure can't be sold. A census would destroy the entire day's production. A fair sample of, say, 100 batteries estimates the average lifetime while leaving 49,900 to ship.
Take each grade in proportion to its share of 2,000, sampling fraction 802000 = 0.04. Freshmen 600·0.04 = 24; sophomores 500·0.04 = 20; juniors 450·0.04 = 18; seniors 18. Check: 24 + 20 + 18 + 18 = 80. ✓
This is voluntary-response bias. Only listeners who felt strongly enough to call took part, and angry people call more often than satisfied ones — so the sample over-represents opposition. The 9,000 figure makes the result precise but not fair: a big biased sample is still biased. A true picture needs a random sample of the whole public.
This is systematic sampling (every k = 10th after the random start 7). It usually spreads the sample nicely. It goes wrong if the list has a hidden rhythm matching k — e.g. if the list runs manager, then nine staff, repeating every 10, then every pick lands on the same role, and the sample misrepresents the office.
Pollster A, despite the far smaller sample. A's method is fair — its 30 estimates wobble around the truth and would tighten if A asked more. B's 300 are drawn from a frame skewed toward the wealthy, so B reports a precisely too-high average. Size cannot cure bias; fairness can.
Six questions to lock it in. Tap the answer you think is right.
This opening lesson of Stage 34 establishes the language of data (population, individual, sample of sizes N and n), the census-versus-sampling trade-off, and the three sampling methods students must distinguish: simple random, stratified, and systematic. It addresses CCSS HSS-IC.A.1 — understanding statistics as a process for making inferences about a population from a random sample — and HSS-IC.B.3 — recognizing the purposes of and differences among sample surveys, experiments, and observational studies, and the central role of randomization. The recurring red thread is that a larger sample does not correct a biased method; the cure for bias is fair, random selection. The next lesson, 34.2, turns the collected data into pictures.