How one sample tells us about everyone.
Up to now we've just described the 200 students in front of us. Starting today, we want to say something about all students — people we never measured. We can't survey everyone, so we take a sample and trust it to stand in for the whole group. This unit is the bridge from "here's what we found" to "here's what's probably true everywhere." One idea does the heavy lifting: the Central Limit Theorem. It's the engine behind every test in the rest of the course.
We can't measure everyone. So we grab a random, representative sample, compute its mean, and use that to estimate the whole group. The catch: a different sample would give a slightly different mean. The Central Limit Theorem tells us exactly how much those means bounce around — and that's what makes guessing about a whole population a science, not a shot in the dark. Here's the cast of characters we'll meet:
| Term | Symbol | Plain English |
|---|---|---|
| Population | — | Everyone you care about (all college students) |
| Sample | n | The smaller group you actually measure |
| Population mean | μ | The true average you're trying to learn |
| Sample mean | x̄ | The average of your one sample — your estimate of μ |
| Sampling distribution | — | All the possible sample means, gathered into one picture |
| Standard error | SE | How much those sample means spread out |
A small convention that pays off all term: Greek letters describe the population (μ for its mean, σ for its standard deviation) and Latin letters describe your sample (x̄ for its mean, s for its standard deviation). The population values are usually unknown — that's the whole reason we sample.
Probability is just a number from 0 to 1 that says how likely something is. 0 means it never happens, 1 means it always happens, and 0.5 is a coin-flip. Multiply by 100 if you'd rather speak in percent (0.5 = 50%). Two things can never happen: a probability below 0 or above 1. If your arithmetic ever produces one, you made an error.
Where does that number come from? It's the long-run relative frequency — the share of times an event happens if you repeat the situation many, many times. Flip a fair coin 10 times and you might get 7 heads; that's noise. Flip it 1,000 times and you'll be hovering right near 500 — about half. Probability is the value that count settles toward as the run gets long. And notice the move: a probability is a count over a total, exactly the relative-frequency calculation you did building tables back in Unit 2.
In our dataset, 18 of the 200 students are vegan. So the probability of randomly drawing a vegan student is 18 ÷ 200 = 0.09 = 9%. Probability is just a count over a total — the same move you made building frequency tables back in Unit 2.
Nobody can measure every college student in America. It would cost a fortune and take forever. So we take a sample — a manageable slice — as long as it's random (everyone has a fair chance of being picked) and representative (it looks like the group it came from). Then we use the sample mean (x̄) to estimate the population mean (μ). A good sample is a small, honest mirror of the whole.
The word "random" is doing real work. Random selection is what protects you from quietly stacking the deck — surveying only your friends, only morning people, only students who answer email. A sample of 50 chosen at random beats a sample of 5,000 that all came from one dorm. How you pick matters more than how many you pick.
This is worth saying twice because it costs people exams and costs companies millions: a big sample that was gathered the wrong way is still biased. More people just gives you a more precise wrong answer. Everything in this unit assumes the sample was drawn fairly to begin with.
Here's the move that unlocks everything. Imagine you don't take one sample — you take many, each of the same size n. For each one, you write down only its mean. Sample 1's mean is 20.3. Sample 2's is 19.7. Sample 3's is 20.1. Keep going for thousands of samples.
Those means pile up into their own distribution — the sampling distribution of the mean. Read that name carefully: it is a distribution of averages, not of individual people. Each dot in it is one whole sample's mean. This is the single most-confused idea in the course, so hold onto the picture: you're drawing a histogram where every value plotted is itself an average of n scores.
And it behaves in a beautifully predictable way. Averages don't swing as wildly as individuals — one unusually tall person barely budges the average height of a roomful of people. So the sampling distribution is narrower than the population, and it's centered right on μ. That predictability is what the next idea names and explains.
This is the star of the whole course. The Central Limit Theorem (CLT) says: as the sample size n grows, the sampling distribution of the mean approaches a normal distribution — no matter what shape the population has. Spelled out, the distribution of sample means is:
That second point is the payoff. The population can be skewed, jagged, weird — the sample means still come out bell-shaped. That's why we can use the normal curve (from Unit 5) for almost any variable. A common rule of thumb: by about n = 30 the sampling distribution is close enough to normal to use, and the more skewed the population, the larger n you'd want. Here's that bell — slide the controls to see how a mean and spread shape it:
Watch what the SD slider does: a small SD gives a tall, tight bell; a big SD gives a wide, flat one. When you picture the sampling distribution of the mean, imagine that bell getting narrower and taller as your sample grows — that's the next idea, and here it is in three pictures:
All three bells sit over the same center, μ — bigger samples don't move the target, they just tighten the aim. The width keeps shrinking, and that width has a name and a formula.
The spread of that distribution of means has a name: the standard error (SE). It's just the standard deviation of the sample means — how far a typical sample mean lands from μ. And there's a clean formula:
SE = σ ÷ √n. Read it slowly. A bigger n (a bigger sample) makes the bottom bigger, which makes the standard error smaller. A smaller standard error means the sample means cluster tighter around the truth — so your estimate is more reliable. Big samples give steady answers.
Notice the √n, not n. That square root is why precision comes at a rising price: to cut the standard error in half you don't double the sample, you have to quadruple it. Going from n = 25 to n = 100 (4×) only halves the spread. The first hundred people you measure buy you far more precision than the second hundred — a real budgeting fact for anyone who runs studies. Let's watch the formula bite, holding σ fixed at 10:
Open the SJNY Student Life dataset. We want the standard error of stress (column J). We already know stress has SD = 5.90 across all n = 200 students. So:
SE = 5.90 ÷ √200 = 5.90 ÷ 14.14 = 0.42
What does 0.42 mean? If we took sample after sample of 200 students, their mean stress scores would typically land within about 0.42 of the true average. That's a tight, trustworthy estimate — because n is big.
And here's the CLT flexing: our commute_minutes (column W) is right-skewed — a few students have brutal 90-minute commutes that stretch it out. But take samples from it and the means still come out normal. The lopsided shape of the raw data doesn't matter. The means behave.
Remember the z-score from Unit 5: z = (X − μ) ÷ σ — how many standard deviations a score sits from the mean. For a sample mean, swap in the same pieces from the sampling distribution: the center is still μ, but the spread is now the standard error. So z = (x̄ − μ) ÷ SE. Because the CLT made the sampling distribution normal, that z drops straight onto the normal curve you already know — and that single substitution is the whole machinery behind every hypothesis test in Block 4.
Paste any list of numbers below (or load class data) to see its mean and standard deviation. Then try the formula by hand: take the SD it reports, pick a sample size n, and compute SE = SD ÷ √n. The bigger the n you imagine, the smaller that standard error gets.
Mean, median, mode, range, variance, standard deviation, and quartiles for one list of numbers — the SD is the σ you'll feed into SE = σ ÷ √n.
This unit isn't abstract — it's how every honest claim about a large group gets made. When researchers report that, say, only a minority of Americans have ever met a vegan, or that public concern for farmed-animal welfare is rising, they did not ask all 330 million people. They drew a random, representative sample of perhaps 1,000–2,000 and let the Central Limit Theorem carry the inference from the few to the many. The standard error is what lets them attach a margin to the claim instead of just guessing.
The same logic powers the studies behind this whole site's running thread. A finding like "meat and dairy use 83% of farmland but provide only 18% of calories" (Poore & Nemecek, 2018) rests on a sample of roughly 38,000 farms standing in for global agriculture — not a census of every farm on Earth. Whenever you read a number about a population too big to count, a sampling distribution is quietly doing the work underneath.
You've seen a poll say "1,500 adults, margin of error ±3%." Why 1,500? Because of everything in this unit. A poll of 1,500 people has a smaller standard error than a poll of 100 — bigger n, smaller SE, tighter estimate. This is the law of large numbers: as your sample grows, its average settles down and stops jumping around.
Flip a coin 10 times and you might get 7 heads (70%). Flip it 1,000 times and you'll be hovering right near 50%. The average calms down as n grows. That's why a national poll of 1,500 beats a quick survey of your 10 friends — and why "we asked more people" really does mean "trust us more" (as long as the asking was random).
You don't compute a standard error by hand — Sheets does it. Open the SJNY Student Life dataset and try the standard-error formula on stress (column J):
=STDEV(J2:J201)/SQRT(COUNT(J2:J201)) → the standard error of stress ≈ 0.42You can also watch the Central Limit Theorem happen with two formulas. =RANDBETWEEN draws a random whole number; wrap a batch of draws in =AVERAGE and you've simulated one sample mean:
=AVERAGE(RANDBETWEEN(1,6), RANDBETWEEN(1,6), RANDBETWEEN(1,6), RANDBETWEEN(1,6))
→ the mean of four random dice rolls — one dot in the sampling distributionFill that formula down 500 rows and you have 500 sample means. Then graph them with Insert ▸ Chart ▸ Histogram: even though a single die roll is flat (uniform — every face equally likely), the averages pile into a bell. That's the CLT, live, in your own spreadsheet. Add more dice to each average and watch the bell get narrower — you've just demonstrated SE = σ ÷ √n by hand.
Standard error is not the standard deviation — SD describes how spread out people are; SE describes how spread out sample means are, and SE is always smaller because you divide by √n (not n). Bigger n shrinks the standard error, never the other way around. The CLT is about means, not individuals — individual stress scores aren't bell-shaped just because the means are. And a bad sample stays bad: the CLT can't fix a sample that wasn't random or representative. Garbage in, garbage out.
Ten quick true/false items with instant explanations. (Different from your Canvas quiz — real practice.)