Populations and Samples

In Statistical Thinking we said your data is one noisy slice of a larger reality. This page gives that idea its proper vocabulary — population and sample, parameter and statistic — and makes one arrow precise: you compute a statistic from a sample to estimate a parameter of a population you can't fully see. Almost every method in the rest of the course is a tool for drawing that arrow honestly.

The distinction sounds almost too simple to dwell on, yet confusing these four words is behind a huge share of real analytical mistakes: reporting a sample average as if it were the population's, expecting a bigger sample to "change" the answer, or forgetting that a statistic wobbles every time you collect new data. Let's pin them down.

Population vs. sample

The population is the entire collection of units you actually care about — every customer, every transaction this design could ever generate, every wafer the machine could ever produce. The sample is the subset you managed to observe and put in a DataFrame.

The loop is the whole idea: the parameter lives in the population (top right), you can't see it directly, so you draw a sample, compute a statistic, and use that statistic to estimate the parameter. The dotted arrows are inference — uncertain by nature. The solid arrows are things you actually do (sampling and computing), which are concrete and exact.

Why we sample instead of taking a census

A census measures every unit in the population. When you can do one cheaply, do it — there's no uncertainty to estimate. But for most data-science questions a census is impossible, impractical, or pointless:

• Cost and time. Surveying all 4 million users, or load-testing every possible session, is wildly expensive when 2,000 well-chosen observations answer the question.
• The population is infinite or future-facing. You literally cannot measure "all sessions this design will ever serve" — most of them haven't happened yet.
• Measurement destroys the unit. A factory testing battery lifetime to failure can't ship the batteries it tested. Destructive testing forces sampling.
• A good sample is enough. This is the punchline of the whole course: a modest, representative sample estimates a population parameter with quantifiable precision. You don't need everything — you need enough, collected well.

Parameters vs. statistics

This is the distinction to burn into memory.

• A parameter is a number that describes the population. It's fixed (the process has some true value) but unknown — you can't see it. Parameters get Greek letters.
• A statistic is a number you compute from the sample. It's known (you can calculate it) but varies — a different sample gives a different value. Statistics get Latin letters or "hats."

A clean way to keep them straight: Greek = the truth you're after; Latin/hat = your best guess from data. The statistic is the estimator; the parameter is the estimand. You use x̄ to estimate μ, p̂ to estimate p, and so on.

Seeing it in code: a known population

Here's the trick that makes this concrete. In the real world the population is hidden, so you never get to check your estimate. But in code we can play god: define a population with known parameters, then draw samples and watch the statistics estimate the (now visible) truth.

The estimate misses by a little — that miss is estimation error, and it's unavoidable from a finite sample. Crucially, the error isn't a mistake you made; it's the price of not seeing the whole population.

A statistic is a random variable

Here's the idea that unlocks everything later. Because a statistic depends on which sample you happened to draw, it changes from sample to sample — which means a statistic is itself a random variable with its own distribution. The sample mean isn't a fixed number; it's a number that would have come out differently with a different sample.

Two things to notice, because they preview the next several pages. First, the sample means pile up around the true μ — the statistic is "aimed" at the parameter (it's unbiased). Second, they have a spread of their own. That spread — how much an estimate wobbles sample-to-sample — is the standard error, and the whole distribution of a statistic across samples is a sampling distribution. We devote entire pages to them: Sampling Distributions and Standard Error.

QuestionSelect one

You compute the mean salary of a random sample of 200 employees and get $72,400. A teammate says "so the average salary at the company is $72,400." What's the precise issue?

Nothing — the sample mean is the company average

The sample is too small for any conclusion

$72,400 is the sample statistic $\bar{x}$, an estimate of the unknown population parameter $\mu$; the company's true average is near it but uncertain

The number is wrong because salaries aren't normally distributed

Estimates have error — and it shrinks as n grows

Estimation error is unavoidable, but it's not uncontrollable. The most important lever is sample size: as n grows, your statistic homes in on the parameter. Watch the sample mean converge.

The error trends toward zero as n grows (it's jagged because each sample is still random, but the envelope shrinks). A key, often-missed detail: precision improves with √n, not n. To halve your error you need roughly four times the data. That diminishing return is exactly why a national poll uses ~1,500 people, not 1,500,000 — past a point, more data buys very little extra precision. We'll quantify this with the standard error formula σ / √n in Standard Error.

QuestionSelect one

You increase your survey from 500 to 5,000 respondents. Which statement is correct?

The population proportion $p$ will move closer to your estimate

The population proportion $p$ is unchanged; your estimate $\hat{p}$ becomes more precise (less sample-to-sample variability)

Both $p$ and $\hat{p}$ stay exactly the same

The estimate's precision improves 10x because the sample grew 10x

Practice the population–sample loop

Check your understanding

QuestionSelect one

Which of these is a parameter (as opposed to a statistic)?

The mean of the 1,200 rows in your loaded DataFrame

The true average lifetime of every battery this factory will ever produce

The proportion of survey respondents who answered "yes"

The standard deviation of last week's recorded sensor readings

QuestionSelect one

Why is it usually impossible, even in principle, to compute a population parameter directly for a data-science question like "does this checkout flow convert well"?

Because parameters are always irrational numbers

Because Pandas can't hold that many rows

Because the population is conceptual and partly in the future — it includes all sessions the flow could generate, most of which haven't happened

Because measurement always changes the parameter's value

QuestionSelect one

A statistic like the sample mean is best described as:

A fixed constant that's the same for any sample from the population

A random variable: its value depends on which sample you drew, so it has a distribution across possible samples

An exact copy of the population parameter it estimates

A value that only exists once you've measured the entire population

QuestionSelect one

Your estimate of a population mean isn't precise enough. You quadruple the sample size from 250 to 1,000. Roughly how much does the typical estimation error shrink, and why?

It shrinks to zero, because 1,000 is large enough to equal the population

It shrinks by a factor of 4, in proportion to the sample size

It shrinks by about a factor of 2, because precision scales with $\sqrt{n}$ and $\sqrt{4} = 2$

It doesn't change, because the population mean is fixed

QuestionSelect one

Pick the statement that correctly uses the notation.

$\bar{x}$ is the population mean and $\mu$ is the sample mean

$\hat{p}$ is the sample proportion and $p$ is the population proportion it estimates

$\sigma$ is computed from your sample and $s$ is the unknown population value

$N$ is your sample size and $n$ is the population size

Next, Sampling and Bias tackles the part this page assumed away — getting a sample that actually represents the population. Then Standard Error quantifies how much an estimate wobbles, and Confidence Intervals turns that wobble into an honest range for the parameter you can't see.