Module 4: Samples · Lesson 4.3 · about 25 min
The central limit theorem
Averages of many independent values form a bell curve centred on the true mean, with spread σ/√n, whatever the raw data looks like. The theorem that makes confidence intervals and tests on averages work, and its limits.
Module 4 · about 25 minutes · school maths. You’ll be able to state the central limit theorem, predict the spread of a sample mean, and know when the theorem can be trusted.
In everyday terms: one dice roll is equally likely to be anything from 1 to 6, flat as a table. Average 30 rolls and you almost always get something near 3.5; the averages pile up in a bell. Highs and lows cancel out.
1. Why averages behave
Individual order values are lopsided: lots of small orders, a few huge ones. Yet the average of 100 random orders behaves beautifully: it lands near the true mean, its possible values form a bell curve, and its spread is predictable. That is the central limit theorem (CLT), probably the most useful result in statistics.
Try each population with n = 1, then 5, then 30. The left plot never changes, but the sample means on the right turn into a bell and narrow as n grows. Even “bought or not” (just 0s and 1s) gives a bell for its averages, which are conversion rates.
2. The theorem
Take random samples of size from any population with mean and standard deviation . For large enough , the sample mean is approximately normal:
Three parts:
- Centre: sample means are centred on the true mean. No systematic error (for a random sample).
- Spread: their SD is , the standard error (next lesson).
- Shape: approximately normal, whatever the population’s shape.
The same holds for totals (sum = n × mean): daily revenue as the total of many orders is close to normal even though orders are not.
3. How large is “large enough”?
- Roughly symmetric data: n of 10–20 is plenty.
- Moderately skewed (most business measures): about 30–50.
- Very skewed or with huge outliers (revenue per visitor, where most are $0 and a few spend thousands): hundreds or thousands. Use the bootstrap (Lesson 4.5) to check.
- Proportions: the rule from Lesson 3.2, and both at least 10.
4. When it fails
- Not independent: 100 orders from 10 customers aren’t 100 independent values. Spread will be bigger than . Analyse at the customer level, or use methods for clustered data.
- Infinite or huge variance: some heavy-tailed data (viral content, insurance catastrophes) never settles.
- Medians and percentiles follow their own rules. The bootstrap handles them.
- Population mean μ, SD σ
- The centre and spread of single values.
- Sample mean x̄ of n values
- One number from one sample.
- x̄ ≈ Normal(μ, σ/√n)
- If you took many samples, their averages would form this bell.
- Physical meaning
- In an average, the lucky highs and unlucky lows cancel each other more and more as n grows. Quadruple the sample and the wobble of the average halves. This is why a daily average is steadier than any single sale.
“The CLT makes the data normal.” It does not change your raw data at all: order values stay skewed. It says the average of many of them is close to normal, which is what intervals and tests use.
Watch it happen, one step at a time:
Maths and principle
Principle: in a sum of many independent pieces, the random highs and lows of the pieces mostly cancel, and what’s left has the bell shape no matter what the pieces looked like.
Why σ/√n: for independent values, variances add (Lesson 3.1). The sum of values has variance ; dividing by to get the mean divides the variance by :
Analogy: one die is flat (each face equally likely); the average of 10 dice is almost always between 2.5 and 4.5, bunched around 3.5. Lesson 3.1’s dice simulation showed the start of this.
5. Worked example
Made-up numbers. A café’s single transactions have mean $8.40 and SD $6.00 (skewed: most are a coffee, some are big group orders). A typical day has 400 transactions.
- Daily average transaction: approximately normal with mean $8.40 and SD 6 ÷ √400 = $0.30.
- About 95 % of days have an average between 8.40 ± 2 × 0.30 = $7.80 and $9.00.
- Daily revenue (400 × average): mean $3,360, SD 400 × 0.30 = $120; 95 % of days between about $3,120 and $3,600.
So a $3,050 day is more than 2.5 SDs low: worth asking why. A $3,250 day is ordinary noise.
6. Practice
- Delivery times have mean 35 and SD 12 minutes. What are the mean and SD of the average of 36 deliveries?
- Why does a mean of 25 values from a very skewed distribution still look skewed?
- A conversion rate is measured on 10,000 visitors with p ≈ 0.03. Can you use the normal approximation?
Answers
- Mean 35; SD 12 ÷ 6 = 2 minutes.
- 25 isn’t large enough to wash out strong skew; the CLT is approximate, and how fast it works depends on the shape.
- Yes: np = 300 and n(1 − p) = 9,700, both well above 10.
Try it on your own data: the bootstrap test tool shows the sampling distribution of the mean (or median) of your own data, so you can see whether it is close to normal.
New words: central limit theorem, sampling distribution of the mean, standard error, independence.
- Averages of n independent values have SD σ/√n.
- Their distribution is close to normal for large n, whatever the shape of single values.
- Very skewed data needs a larger n before the bell appears.
- The CLT fails for dependent data and for extremely heavy tails.
Check yourself
Sources
- OpenIntro Statistics, 4th edition (Diez, Cetinkaya-Rundel, Barr), Section 5.1 Point estimates and sampling variability https://www.openintro.org/book/os/
- Blitzstein, J. and Hwang, J. (2019), Introduction to Probability, 2nd edition, Section 10.3 Central limit theorem https://projects.iq.harvard.edu/stat110
Your ticks stay in this browser. Save them as a file on the course home.
Updated 2026-11-22. Numbers in the examples are made up to show the idea unless a source is given. Simulations run in your browser; nothing you type is sent anywhere.