Interactive statistics

56 free interactive lessons: from mean and median to A/B tests and Bayes. Drag the charts and build intuition. No sign-up.

Module 4: From sample to world

An estimate is a random variable (the key idea of inference)

This may be the most important lesson of the whole course — keep it in mind through all the following ones. All we have is a sample, and from it we ESTIMATE something: a mean, a share, a difference. The key insight: any such estimate is itself a random variable. Take another sample — you get another number. And this "wandering" has its own distribution, on which all of statistical inference rests.

Estimate:
truth (mean) = 11.3population (fixed truth)distribution of the “mean” estimate across samples
Samples: 8n: 20center of estimates ≈ 10.9 (≈ the truth)

Whichever estimate you pick — every sample has its own, i.e. it is itself random. Its spread falls as n grows. That is the sampling distribution — the object the CLT, intervals and hypothesis testing all stand on.

Let's separate two things that are easy to confuse. The truth (the estimand) is a number about the whole population: the true mean μ, the true share p. It is single and fixed, but unknown to us. The estimate (the estimator) is what we compute from a sample: x̄, p̂. It is known to us, but changes from sample to sample. The yellow line is the truth (one); the teal histogram below is how the estimate dances.

What it means

This frame instantly changes how you read any number. "Conversion is 12%" is not the truth but one realization of a random estimate; the right question is "how wide is its sampling distribution?". Intervals and significance testing grow naturally from here.

And this is also why sample size matters so much: it directly squeezes the estimate's sampling distribution, i.e. reduces its random wandering around the truth.

Where it shows up

Polls, ratings, quality measurements, product metrics — everywhere you hold only a sample-based estimate. A professional automatically keeps its sampling distribution in mind and never confuses "the number we got" with "the truth".

That is why serious reports show not a bare point but an estimate with an interval — an honest admission that it is a random variable, not a verdict.

Definitions
Estimand
the true quantity about the whole population (μ, p). Fixed but unknown.
Estimator
what we compute from the sample (x̄, p̂). Known but random — changes from sample to sample.
Sampling distribution
the estimate's distribution over all possible samples. Its center is the truth; its width is the estimate's uncertainty.
Unbiasedness
the property of an estimate being equal to the truth on average (the sampling distribution is centered on the estimand).
When the method lies (assumptions)

The formula for the standard error, σ/√n (meet it properly in the next lesson), assumes independent observations. If rows are clustered (several visits by one person, cities, ad campaigns), the effective n is smaller than the row count — the real error is larger than the formula says.

The standard error describes only random scatter. Systematic error — an unrepresentative sample, a biased measurement — is not cured by √n: as n grows you converge ever more confidently to the wrong answer.

Deep dive: the math and the mechanism (optional)

Формально оценка — это функция от случайной выборки, поэтому она сама случайная величина. У неё есть математическое ожидание (куда она целится в среднем) и дисперсия (как сильно гуляет). Несмещённость — это «ожидание оценки = оцениваемое»; эффективность — «при той же выборке дисперсия меньше».

Дальше мы увидим, что для среднего это распределение имеет ширину σ/√n и стремится к нормальному (ЦПТ) — но сама идея «оценка случайна и имеет распределение» верна для любой статистики, даже когда формулы нет (тогда выручает бутстреп).

Next →
Enjoying the materials?

This site is built by the DataSlice Telegram channel: data analytics in plain words — real cases, metrics, careers. The channel is in Russian.

Open the Telegram channel