Join · scan
OTTO-INF1
ottolearn.org/join
Module 4 · Statistical Inference · L-INF-1
A letter lands on Maya's desk: "You surveyed 300 of 48,000. A different 300 would've given a different average. Your 6.8 is one roll of the dice — why should anyone trust it?" The reader is right that a different 300 gives a different average — and today we meet the mathematics of exactly how different: the distribution of the averages themselves. It's the question Maya has carried since her very first survey, finally becoming answerable.
Live session · join with code OTTO-INF1 · ottolearn.org/join
Same rules as always: vote before I explain, every poll is anonymous, and changing your mind after discussion is the goal.
Today · one letter, five answers
The sample mean is a random variable — the sampling distribution
Center \(\mu_{\bar{x}} = \mu\) and standard error \(\sigma/\sqrt{n}\)
The Central Limit Theorem — and its license conditions
The \(\sqrt{n}\) law and its diminishing returns
Last class's pipeline, upgraded: standardize with the SE
Meet the distribution of averages, find its center and its spread, learn the theorem that fixes its shape, see what sample size buys — then compute how rare any average is. Every step starts with your vote.
Recall · last class · 45 seconds
Warm-up · commit first · 60 seconds · anonymous
Hold your answer — it is today's lesson. By the end of class you'll estimate the typical sampling variation of an average from 300 students, and the number is smaller than almost everyone's first instinct.
Core concept 1 · C1
The skeptic's objection lives one level above anything we've drawn before. Keep three objects apart: the population (all values), one sample (the values Maya got), and — new today — the sampling distribution: the distribution of the averages from every possible sample.
All 48,000 students' sleep hours. Its histogram is fixed but never fully seen; its mean \(\mu\) and standard deviation \(\sigma\) are the unknown truth Maya is after.
\(N = 48{,}000\ \cdot\ \mu,\ \sigma\)Maya's 300 responses — the data she actually holds, with \(\bar{x} = 6.8\) h. A different random 300 would be a different level-2 object with a different \(\bar{x}\).
\(n = 300\ \cdot\ \bar{x} = 6.8\)Imagine every possible sample of 300, each boiled down to its average — then the histogram of all those \(\bar{x}\)'s. Never directly observed; it's the theoretical object that describes how \(\bar{x}\) varies around \(\mu\).
\(\mu_{\bar{x}},\ \sigma_{\bar{x}}\)The skeptic is asking a level-3 question: "how much do the \(\bar{x}\)'s scatter?" — and level 3 is precisely the thing built to answer it. Watch all three levels side by side:
Quick check · the name trap
See it · C1
Predict first: I'll press Draw 1 sample five times. Which panels change and which stay fixed? Then I'll draw 100 samples from this right-skewed population — will the pile of averages in panel ③ be skewed too, and how wide will it be?
Commit first · today's summit
Core concept 2 · C2
The level-3 object has a center and a spread, and both come out of formulas — no simulation needed. The familiar formulas assume independent observations with finite \(\sigma\); sampling without replacement needs a finite-population correction when the sampled fraction is not small.
Under the stated random-sampling model, the pile of \(\bar{x}\)'s centers on the population mean: \(\mu_{\bar{x}} = \mu\). Sample means don't systematically over- or under-shoot — \(\bar{x}\) is an unbiased estimator.
\(\mu_{\bar{x}} = \mu\)The typical distance between a sample average and \(\mu\) is the standard error \(\text{SE} = \sigma/\sqrt{n}\) for independent observations. Maya observed \(s = 1.3\) h, so her estimated SE is \(s/\sqrt{300} \approx 0.075\) h.
\(\sigma_{\bar{x}} = \dfrac{\sigma}{\sqrt{n}}\)Individual sleep values have sample SD \(s = 1.3\) h; the estimated SD of averages of 300 is about \(0.075\) h. A standard deviation is a yardstick, not a guaranteed \(\pm\) range.
\(\sigma \neq \sigma_{\bar{x}}\)The warm-up, quantified: under an approximately normal plug-in model, roughly 95% of comparable sample means would fall within about \(2\widehat{\text{SE}} \approx 0.15\) h of \(\mu\). That is a model-based frequency, not a guarantee about every rerun. Watch the two yardsticks live:
See it · C2
Predict first: as I drag \(n\) from 9 to 100, the individuals' curve and the averages' curve — which moves, which stays? And does the averages' curve ever get wider?
Practice · the yardstick for an average · fresh numbers on demand
Core concept 3 · C3
Center and spread are settled. The last question about the pile of averages is its shape — because only a known shape lets us read probabilities off a table. The answer is one of the most remarkable theorems in statistics.
For independent observations from a population with finite \(\sigma\), the pile of averages approaches a bell as \(n\) grows. The population may be skewed or bimodal, but there is no universal sample size that makes every case adequate.
\(\bar{x} \approx N\!\left(\mu,\ \sigma^2/n\right)\)Independence — supported by the sampling process. Finite \(\sigma\). Shape and size together — \(n = 30\) is only a starting guideline; severe skew, heavy tails, or influential outliers may require much more. A normal population gives an exactly normal pile at any \(n\).
\(n \geq 30\)Your data never turns normal — 50 skewed incomes stay skewed at any \(n\); only the pile of averages goes bell. And no sample size rescues broken independence: a time series or repeated measures on the same people fails at \(n = 20\) and at \(n = 20{,}000\).
This is why the bell from last class is everywhere: even when individual values refuse it, their averages wear it. Watch the bell assemble out of a skewed population:
Quick check · what goes normal?
See it · C3
Predict first: the population is right-skewed. At \(n = 2\), will the pile of averages be skewed or bell-shaped? Then I raise \(n\) to 30 and draw 100 samples — call the shape before it lands.
Practice · is the bell licensed? · fresh numbers on demand
Core concept 4 · C4
The \(\sqrt{n}\) in the denominator is the entire economics of survey design. It explains why Maya sampled 300 — and why she didn't sample 3,000.
Using Maya's \(s = 1.3\) h to estimate \(\sigma\): the estimated SE is \(0.13\) h for \(n = 100\) and \(0.075\) h for \(n = 300\). Every extra independent respondent tightens the average's aim.
\(\sigma_{\bar{x}} = \dfrac{\sigma}{\sqrt{n}}\)To halve the estimated SE you must quadruple \(n\) — doubling only buys a factor of \(\sqrt{2} \approx 1.41\). Going from 300 to 1,200 interviews would reduce it from about 0.075 h to about 0.038 h: four times the work for one extra decimal of calm.
\(\sqrt{4n} = 2\sqrt{n}\)A small SE is what lets a real deviation stand out from sampling noise. The narrower the pile of averages, the harder it is for "we just got an unusual sample" to explain away a surprising \(\bar{x}\) — the idea the last stretch of today runs on.
The skeptic asked "why trust 300 of 48,000?" Here \(n/N = 0.625\%\), so the finite-population correction changes the estimated SE only from about \(0.07506\) h to \(0.07482\) h. Sample size is the main driver here because the sampled fraction is tiny; for a large fraction, \(N\) matters too. See the curve of diminishing returns:
See it · C4
Predict first: where does the curve flatten — before or after \(n = 100\)? Find the point where doubling the sample stops feeling worth it.
Practice · what sample size buys · fresh numbers on demand
Core concept 5 · C5
Everything assembles here. The pile of averages is normal (C3), centered on \(\mu\) (C2), with spread \(\text{SE} = \sigma/\sqrt{n}\) (C2, C4) — so last class's entire z-pipeline works on averages verbatim, with exactly one edit.
Before any z, compute the yardstick that actually fits an average: \(\text{SE} = \sigma/\sqrt{n}\). This step is the whole upgrade — everything after it is old news.
\(\text{SE} = \dfrac{\sigma}{\sqrt{n}}\)Same bridge as last class, new footing: \(z = (\bar{x} - \mu)/\text{SE}\). Divide by \(\sigma\) instead and the z comes out \(\sqrt{n}\) times too small — the trap with a starring role today.
\(z = \dfrac{\bar{x} - \mu}{\text{SE}}\)Read the left-tail entry, complement for "greater than", subtract for "between". If you can read the table for one value, you can read it for an average — the table never knows the difference.
\(P(Z < z)\)Name the object — one value or an average of \(n\)? — and the object picks the denominator. See the pipeline as one picture, then watch me run a full verdict out loud:
I do · watch a full verdict
Watch me decide every step out loud — name the object first (one value, or an average?), then SE, z, read, convert, verdict.
Readers tip Maya off; the plant's spec says fills are normal with \(\mu = 500\) g, \(\sigma = 8\) g. She weighs a random \(n = 64\) boxes: \(\bar{x} = 498\) g. That 498 is an average of 64, not one box — so its yardstick is the SE, not \(\sigma\). The question: \(P(\bar{x} \leq 498)\) if the machine is honest.
\(\text{SE} = 8/\sqrt{64}\) \(= 1\) g. Averages of 64 boxes wander by about a gram — so 498 sits two whole yardsticks below spec. I expect a small area; writing that expectation down is the cheapest error alarm there is.
\(z = \dfrac{498 - 500}{1}\) \(= -2.00\). Had I divided by \(\sigma = 8\) instead: \(z = -0.25\) — an ordinary-looking blip. The wrong denominator doesn't just change a number; it changes the story.
Table: \(P(Z < 2.00) = 0.9772\), so \(P(Z < -2.00)\) \(= 1 - 0.9772 = 0.0228\). If the machine were honest, only about 2 samples in 100 would average this low. That's not proof — but it's printable evidence, and Maya has a story.
The habit: name the object (one value, or an average of \(n\)?), compute the SE, standardize, read and convert, then say the verdict in words.
We do · you finish it
Same tip, smaller sample: you re-run the pipeline and supply the verdict — and watch the sample size flip it.
Suppose Maya had weighed only \(n = 16\) boxes and found the same \(\bar{x} = 498\) g. Re-run the pipeline: same spec (\(\mu = 500\), \(\sigma = 8\)), same question — \(P(\bar{x} \leq 498)\).
Finish it: \(P(\bar{x} \leq 498) = ?\) — and does the verdict survive?
See it · C5
Predict first: which single box in this pipeline is new since last class? Everything else you've already done.
Which tool applies? · mixes earlier classes + today
Practice · the full pipeline, solo · fresh numbers on demand
Now you run the whole pipeline solo — fresh numbers every re-roll, the table entries you need on screen, support at its minimum.
Production task · answer the skeptic
Exit ticket · 60 seconds · anonymous
What Maya files
The objection was correct and quantifiable: \(\bar{x}\) is a random variable. Its distribution — the sampling distribution — centers exactly on the truth (\(\mu_{\bar{x}} = \mu\)), wanders only \(\text{SE} = \sigma/\sqrt{n}\), and is bell-shaped by the Central Limit Theorem once independence and size conditions hold.
For her survey: \(\widehat{\text{SE}} = s/\sqrt{300} = 1.3/\sqrt{300} \approx 0.075\) h. Under the approximate normal plug-in model, about 95% of comparable sample means fall within roughly \(\pm 0.15\) h of \(\mu\). The cereal check showed the pipeline's teeth: the same 498 g is a 2-in-100 alarm at \(n = 64\) and a 1-in-6 shrug at \(n = 16\).
The one upgraded move: name the object first. One value → \(\sigma\). An average of \(n\) → \(\sigma/\sqrt{n}\). A count of trials → the binomial. After that, it's the same table, the same left-tail sentence, the same convert-before-you-report.
Maya can now model how sample averages vary — but "roughly 95% within about ±0.15 h of \(\mu\)" is still a sampling-distribution statement, not an interval calculated from her observed mean. Next class builds the confidence interval around 6.8 and states its confidence level.
Vote counts on today's slides are simulated for rehearsal; live sessions show the room's real votes.