Statistics › Section G · The central limit theorem › Topic G.3
Using the central limit theorem
The CLT gives the recipe; this topic runs it in the kitchen. A checklist for when the theorem applies, the choice between averages and sums, and the reading of rare results - the bridge every later inference topic crosses.
The CLT in action
The two previous topics built the central limit theorem's machinery: averages centred at \(\mu\) with standard error \(\sigma/\sqrt{n}\), sums centred at \(n\mu\) with SD \(\sigma\sqrt{n}\), both approximately normal for large samples. This topic runs that machinery on practical problems: when the theorem applies, which statistic matches which question, and what a surprisingly extreme result should make you suspect. The examples are deliberately ordinary - bakeries, deliveries, monthly totals - because the CLT's point is exactly its ordinariness: the same three steps answer nearly every averaging question in the wild.
The checklist before computing
Four conditions guard every CLT calculation. 1. Random sampling: the data are a chance sample, not a convenience grab. 2. Independence: observations do not influence each other. 3. Sample size: \(n \ge 30\) is the usual rule of thumb, more for very skewed populations. 4. For the calculations in this chapter, known centre and spread: the problem supplies \(\mu\) and \(\sigma\) (later topics estimate \(\sigma\) from the sample). No condition says the population must be normal - that is the theorem's whole gift - but a skewed population does ask for a larger \(n\) before the bell approximation is trustworthy.
Choose the statistic by the question
"How many per day, on average?" is an average question: standardise with \(\frac{\bar{x} - \mu}{\sigma/\sqrt{n}}\). "Can 36 days' sales reach the target?" is a total question: standardise with \(\frac{\sum x - n\mu}{\sigma\sqrt{n}}\). The two are one theory apart - \(\sum X = n\bar{x}\) - but a mixed-up scaling gives a spread that is wrong by a factor of \(n\). Read the question's units first, write the matching z-formula second, and only then reach for the table - reversing the order is where most wrong answers begin.
Worked example: average daily loaves
A bakery sells \(\mu = 200\) loaves a day with \(\sigma = 30\), day to day independent. Over a random 36 days, what is the chance the daily average tops 210? 1. Checklist: random days, independent, \(n = 36\) large enough. 2. Standard error: \(\frac{30}{\sqrt{36}} = 5\) loaves. 3. Standardise and read the upper tail:
4. Check: 210 is 2 standard errors above centre, and the empirical rule puts roughly 2.5% beyond that. 5. Percentiles run backwards: the 95th percentile of the 36-day average is \(200 + 1.645 \times 5 \approx 208.2\) loaves.
Worked example: a month's total loaves
Same bakery: can 36 days of sales total more than 7,500 loaves? 1. Total's mean: \(36 \times 200 = 7200\). 2. Total's SD: \(30 \times 6 = 180\). 3. \(z = \frac{7500-7200}{180} = 1.67\) to two decimals, so
4. Check: divide the sum statement by 36 and it becomes the average statement "\(\bar{x} > 208.3\)" - between the previous example's 210 boundary and the centre, so a probability between 0.0228 and 0.5, which 0.0475 is. The two statistics agree about the world.
Reading a rare result
Suppose a supplier claims a mean delivery time, and your random sample of 50 orders averages 4 standard errors above the claim. Under the claim, a sample mean that far above happens about 3 times in 100,000, since \(P(Z > 4) \approx 0.00003\). Two explanations survive: a staggeringly rare event occurred, or the claimed \(\mu\) is wrong. Statisticians do not accept "miraculous luck" as a working explanation when a simpler one sits available - the claim itself is suspect. This reading is the seed of hypothesis testing, developed in its own chapter: measure how far the sample lands from the claim, in standard errors, and let the distance argue.
The law of large numbers
A companion result explains why bigger samples are worth collecting. By the law of large numbers, a bigger sample's mean \(\bar{x}\) is more likely to sit near the population mean \(\mu\), and in the long run it settles on \(\mu\). A fair die has \(\mu = 3.5\) and \(\sigma \approx 1.71\). Ten rolls might easily average 4.1, but the standard error for 10,000 rolls is \(\frac{1.71}{100} \approx 0.017\), so their average lands within about 0.03 of 3.5 roughly 95% of the time. The two results fit together: the law of large numbers says where \(\bar{x}\) is heading, and the central limit theorem describes the shape and size of its scatter on the way. Neither says a short run of high rolls must be balanced by low ones.
When σ is unknown
Real problems rarely hand over \(\sigma\) wrapped in ribbon. The practical fix, foreshadowed here and developed in the confidence-interval chapter: estimate the spread with the sample standard deviation \(s\), so the standard error becomes \(\frac{s}{\sqrt{n}}\). The substitution adds a little extra uncertainty, and the bell must be widened slightly to compensate - that widened curve is the \(t\) distribution, waiting one chapter ahead. For now, every problem supplies \(\sigma\); the machinery is identical once the estimate arrives, and only the bell's name changes.
Common mistakes to avoid
First, using the CLT on a convenience sample: the theorem assumes chance sampling, and bias survives any amount of averaging. Second, scaling with \(n\) where \(\sqrt{n}\) belongs - means and sums each have their own formula, and swapping them puts the spread off by a factor of \(n\). Third, trusting the bell at \(n = 5\) on heavily skewed data. Fourth, answering a total question with the average's standard error. Fifth, treating a rare result as proof: an extreme sample makes the claim doubtful, but it does not by itself prove the claim false. Sixth, plugging \(\sigma/\sqrt{n}\) in when the question describes single observations, not means - the SE belongs to averages alone.
Key takeaways
Before any CLT calculation, run the checklist: random, independent, large enough \(n\), known \(\mu\) and \(\sigma\) - no normality of the population required. Match the statistic to the question: averages standardise as \(\frac{\bar{x}-\mu}{\sigma/\sqrt{n}}\), sums as \(\frac{\sum x - n\mu}{\sigma\sqrt{n}}\), and \(\sum X = n\bar{x}\) links them. The bakery: average over 210 loaves, 0.0228; total over 7,500, 0.0475 - consistent with each other. The law of large numbers pulls \(\bar{x}\) towards \(\mu\) as \(n\) grows. A sample landing several standard errors from a claim makes the claim, not luck, the suspect - the seed of hypothesis testing. Unknown \(\sigma\) is the confidence-interval chapter's opening problem.
Flashcards
Flashcard
Random, independent, large n, known μ and σ
Test your understanding
Question 1 of 5
Which condition is NOT required for the central limit theorem to apply to a sample mean?
Question 2 of 5
The bakery (μ = 200, σ = 30) samples 36 days. What is P(x̄ > 210)?
Question 3 of 5
Same bakery: what is P(36-day total > 7500 loaves)?
Question 4 of 5
The bakery asks whether 36 days of sales can reach a 7,500-loaf contract. Which statistic answers?
Question 5 of 5
A supplier claims a mean time, but your random sample lands 4 standard errors above the claim. The best reading:
