A spoonful judges the pot — the sample mean estimates μ, more data steadies it, and the bell tells you what is normal.
You cannot measure everyone. To know the average height of every 16-year-old in the country, you do not line up forty million teenagers — you measure a few hundred, take the sample mean x̄, and trust it to speak for the population mean μ you will never see. That trust is the whole of inference: a careful leap from the part you held to the whole you did not. The leap is never certain — one sample always comes out a little different from the next — but it is governed by laws so steady you can put numbers on the wobble. This lesson shows why the leap is safe, when it is not, and how the bell curve turns "probably" into "95% of the time."
Stage 33 asked how likely is tomorrow? Statistics asks the mirror question: the data is already in hand — what does it say about the whole? We almost never get the whole population (every teenager, every light bulb a factory will ever make). So we take a sample of size n, compute a statistic from it, and use that statistic to estimate the matching feature of the population.
The headline pair: the sample mean x̄ estimates the population mean μ. Sip one spoonful of a well-stirred pot and judge the whole pot by it. A sample proportion works the same way: the fraction of your sample that is left-handed estimates the population proportion p.
Five students' heights (cm): 171, 168, 175, 169, 172. The sample mean is 171+168+175+169+1725 = 8555 = 171 cm. Our best single guess for the population mean μ of all students' heights is x̄ = 171 cm. It is a guess, not the truth — but it is the honest one the data supports.
Conclusion: a sample statistic is our best estimate of the matching population value — x̄ → μ, sample proportion → p.
Where does the trust come from? From the most stubborn fact in Stage 33: the relative frequency of an event settles down. Flip a fair coin ten times and you might see 7 heads — frequency 0.7, nowhere near a half. Flip it five hundred times and the running fraction creeps toward 0.5 and stays. The long-run frequency is the probability; that is why a sample frequency is a trustworthy estimate of an unknown probability.
This is the same engine as x̄ → μ, written for a yes/no question. Counting "heads" as 1 and "tails" as 0, the running proportion of heads is just the running mean of those 0s and 1s — so a stabilizing frequency and a stabilizing mean are one phenomenon, not two. (See Stage 33 · Random phenomena and the birth of probability.)
Conclusion: a sample frequency estimates a probability, and the estimate gets steadier the more you observe.
Here is the catch and the cure in one breath. Two honest samples from the same population will hand you two different x̄'s — that is sampling variability, and it is unavoidable. But the wobble is not random in size: the larger the sample, the smaller the wobble. A sample of 4 can land far from the truth; a sample of 64 hardly dares. This is the law of large numbers, felt rather than proved: x̄ closes in on μ as n grows.
Picture the truth as a target. Each sample fires one dot — the sample mean — at it. With small n the dots spray wide around the bullseye; crank n up and the dots tighten into a knot around μ. Notice the dots still center on μ the whole time (a fair method aims true); only their scatter shrinks.
One sample is not the truth; it is an estimate with wobble. And a bigger sample only tightens the wobble — it does not fix a biased method. Measuring heights on a basketball court and asking for ten thousand players just buys you a precise wrong answer: the dots cluster tightly, but around the wrong center, not μ. Size cures variance, not bias (Stage 34.1).
Conclusion: x̄ drifts toward μ as n rises — but only if the sample was fair to begin with.
Measure the heights of thousands of people and a shape appears, the same shape every time: a smooth hill, high in the middle, thinning symmetrically to either side. Most people are near average; a few are quite short or quite tall; almost none are extreme. This is the normal distribution, written N(μ, σ²) — the bell that quietly governs heights, weights, measurement errors, and (as 34.3 hinted) sample means themselves.
Two numbers fix the whole curve. The mean μ says where the peak sits — slide μ and the bell slides along the axis without changing shape. The standard deviation σ says how wide — small σ gives a tall narrow spike (everyone alike); large σ gives a low broad mound (everyone different). The area under the whole curve is always 1, because somebody has every value.
For a smooth curve like the bell, the height at a point is a density, not a chance — no single exact value (a height of exactly 170.0000… cm) has positive probability. What carries probability is area: the chance of landing between a and b is the area under the bell from a to b. This is the same "area = probability" idea from Stage 33's continuous distributions, and the same "area sums to 1" idea you met under a relative-frequency histogram (34.2).
Conclusion: the bell N(μ, σ²) is fixed by two numbers — μ locates it, σ spreads it — and area, not height, is probability.
The bell pays off in three numbers worth memorizing for life. Because every normal curve is the same shape, the fraction of values within a given number of standard deviations of the mean is always the same — no matter what μ and σ happen to be:
| Interval | Fraction inside | Roughly | Outside (both tails) |
|---|---|---|---|
| μ ± 1σ | 0.6827 | ≈ 68% | ≈ 32% |
| μ ± 2σ | 0.9545 | ≈ 95% | ≈ 5% |
| μ ± 3σ | 0.9973 | ≈ 99.7% | ≈ 0.3% |
So about two-thirds of the population is within one σ of average, nineteen in twenty within two σ, and all but three in a thousand within three σ. Anything beyond 3σ is so rare — under 0.3% — that we call it unusual and start to suspect something special is going on. That instinct — "this is too far out to be ordinary chance" — is the seed of statistical judgment that Stage 35 · Logic will sharpen into a formal test.
Adult heights are roughly normal with μ = 170 cm and σ = 8 cm.
• Within 1σ, i.e. 162 to 178 cm: normBetween(162, 178) = 0.6827 — about 68% of people.
• Within 2σ, i.e. 154 to 186 cm: 0.9545 — about 95%.
• Taller than 194 cm (that is μ + 3σ): only 0.00135 — about 1 person in 740. Genuinely unusual.
Conclusion: 68 – 95 – 99.7 within 1, 2, 3 standard deviations; beyond 3σ is rare, not impossible.
| Idea | What it says | The number |
|---|---|---|
| Estimate the whole | the sample speaks for the population | x̄ → μ, prop → p |
| Frequency stabilizes | a frequency settles to a probability | heads → 0.5 |
| Bigger n, less wobble | more data ⇒ steadier estimate (law of large numbers) | spread of x̄ ↓ as n ↑ |
| The bell | N(μ, σ²): μ centers, σ widens; area = probability | total area = 1 |
| The 3σ rule | fraction within 1σ / 2σ / 3σ | 68% / 95% / 99.7% |
And the standing warning: an estimate from a biased method stays wrong however large the sample. Honest randomness first (34.1); then the bell does the rest.
A sample of six bags of flour weighs (kg): 1.02, 0.98, 1.05, 1.00, 0.97, 1.04. Estimate the mean weight μ of the whole production run.
x̄ = (1.02+0.98+1.05+1.00+0.97+1.04)/6 = 6.06/6 = 1.01 kg. Our best estimate of the population mean is μ ≈ 1.01 kg — a guess, not the truth, but the one the sample supports.
You flip a coin 20 times and get 13 heads; a friend flips it 2000 times and gets 1012 heads. Whose head-frequency is the more trustworthy estimate of the true probability, and why?
The friend's. Frequencies: yours = 13/20 = 0.65; the friend's = 1012/2000 = 0.506. A frequency stabilizes as the number of trials grows, so the larger run is far steadier and closer to the true 0.5. The small run's 0.65 is just sampling wobble.
Two pollsters estimate the same population mean. One samples n = 100 people, the other n = 2500. Both use fair random sampling. Which estimate has the smaller wobble, and roughly how much smaller is its spread?
The n = 2500 estimate wobbles less — more data steadies x̄. The spread of a sample mean shrinks like 1/√n, so going from 100 to 2500 (a 25× larger sample) cuts the spread by √25 = 5×. (Bigger n; tighter dots — exactly what the simulator shows.)
Test scores are normal with μ = 500 and σ = 100. Using the 68–95–99.7 rule, what fraction of scores fall between 400 and 600? Between 300 and 700?
400 to 600 is μ ± 1σ → about 68% (exactly 0.6827). 300 to 700 is μ ± 2σ → about 95% (exactly 0.9545).
In the same N(500, 100²) distribution, a score of 820 shows up. How many standard deviations above the mean is it, and should you be surprised?
(820 − 500)/100 = 3.2 σ above the mean. Beyond 3σ is under 0.3% of values, so yes — unusual. It is not impossible, but it is rare enough to make you look twice (a brilliant student, or a scoring error?).
A company surveys only people who phoned its customer-help line and finds their average satisfaction is high. They sample 50,000 callers. Is "average satisfaction is high" a safe estimate for all customers? Explain.
No. The sampling frame — only callers — is biased; people who call may differ systematically from everyone else. A huge sample makes the estimate precise, but it is precisely estimating the wrong population. Size cures wobble, not bias.
Six questions to lock it in. Tap the answer you think is right.
This lesson introduces statistical inference: using a random sample to estimate a population parameter (HSS-IC.A.1), recognizing that a model's fit is judged through sampling (HSS-IC.A.2), and that estimates of a population mean or proportion carry sampling variability that shrinks with sample size (HSS-IC.B.4). The closing sections develop the normal distribution and the empirical 68–95–99.7 rule (HSS-ID.A.4). Throughout, every probability and spread is computed by a shared math engine — the 68/95/99.7 figures come from GX.normBetween (0.6827 / 0.9545 / 0.9973), and the simulator's tightening scatter is the standard deviation of genuine resampled means — so the pictures never claim more than the arithmetic earns. The standing caution that a large sample cannot rescue a biased method connects back to Stage 34.1, and the "too far out to be ordinary" instinct points forward to the formal reasoning of Stage 35 · Logic.