The result that makes the rest of the course possible. Chapter 6's normal machinery applies only to quantities that happen to be bell shaped, but the central limit theorem says that if random samples of size n are drawn from any population at all, then as n increases the distribution of the sample means tends toward a normal distribution. That distribution has the same mean as the original, and a standard deviation equal to the original standard deviation divided by the square root of the sample size, a quantity called the standard error of the mean. Two consequences follow. Sample means are far less variable than individual values, and they become less variable in proportion to the square root of the sample size rather than to the sample size itself. The sample size n counts the values averaged together to form one mean, not the number of times the sampling is repeated.
Subject: Statistics · 70 slides · symbolic lesson
Open the interactive version of this deck
Title
Statistics · Chapter 7 — The Central Limit Theorem
Average enough values from any population and the average is normal, with a spread you can compute
Objectives
Chapter 6 gave you the normal curve, but only for quantities that happen to be bell-shaped. Most real populations are not: incomes, waiting times and pocket change are all skewed. This section shows that averages of them are normal anyway, and tells you exactly which normal curve they follow.
One idea runs through all five: a sample mean is itself a random variable, with its own distribution, and that distribution is narrower than the population's. Once you know its centre and its spread, every calculation is the Chapter 6 calculation you already know, with one number changed.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 365-370 — the chapter introduction and the section this deck covers
Warm-up
Two skills from earlier chapters carry this whole section. Recall both on small cases now.
Discussion prompt
First, five dice show 2, 2, 3, 4 and 6: what is their mean? Second, if X is normal with mean 50 and standard deviation 10, what is the z-score of 65, and roughly what share of values lie between 40 and 60?
If the z-score felt rusty, remember what it measures: distance from the mean counted in standard deviations. This section keeps that meaning and changes only which standard deviation you count in.
Hold on to the first answer: the book's own collaborative exercise opens with exactly that roll of five dice, and 3.4 is one value of a new random variable. Hold on to the second too: every probability in this deck ends as a z-score on a normal curve.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 366 — the five-dice roll 2, 2, 3, 4, 6 with mean 3.4
Section
Part 1
Concept
The book starts with a class experiment. Some of you roll one die, some two, some five, some ten. Each time you roll, you record the faces and compute their mean. Here is one roll of five dice from the book.
Figure (svg): Five dice showing 2, 2, 3, 4 and 6. Under them the arithmetic 2 plus 2 plus 3 plus 4 plus 6, over 5, equals 17 over 5, equals 3.4, labelled one value of the sample mean for n equals 5. An arrow points to x-bar equals 3.4.
\[ \bar{x} = \frac{2 + 2 + 3 + 4 + 6}{5} = 3.4 \]
The number 3.4 is one sample mean for a sample of size five. Roll the same five dice again and you get different faces and a different mean, perhaps 4.2 or 2.8. So the sample mean is not a fixed number: it varies from sample to sample, and it has a distribution of its own.
That distribution is what this chapter is about. Each person in the experiment rolls ten times, collecting ten means, and the class pools its means into one histogram for each number of dice. The question is what shape those histograms take, and the answer turns out not to depend on the dice at all.
Notice the notation: a capital X-bar names the random variable, the mean of a sample not yet taken, and a lower-case x-bar names one value it took, like 3.4.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 365-366 — the collaborative exercise
Prediction
One die is flat: each face from 1 to 6 comes up one time in six, so a histogram of single rolls is six equal bars. Now imagine the class histogram of means of ten dice. Commit to a prediction before you see it.
Predict first
A histogram of many means of ten fair dice will look like which of these?
Correct: a narrow bell centred at 3.5
Why: Averaging ten faces lets highs and lows cancel, so means near 3.5 are common and means near 1 or 6 need ten extreme faces at once, which almost never happens. The result is a bell, narrow and centred at the die's own mean of 3.5.
The flat option is the tempting one if you picture the population and forget the averaging. To get a mean of 1 with ten dice you need ten ones in a row, a chance of one in about 60 million; to get a mean near 3.5 you need only a balanced mix, and there are millions of balanced mixes.
\[ P(\bar{x} = 1) = \left(\frac{1}{6}\right)^{10} \approx 1.65 \times 10^{-8} \]
Counting the ways is the whole reason for the bell: middle values of the mean can be made in many ways, extreme values in very few.
Picture it
Here are the class histograms, drawn exactly rather than simulated. Each row is the distribution of the mean for a given number of dice, computed by listing every possible roll, and all four rows share one height scale.
Figure (svg): Four rows of bars on one horizontal scale from 1 to 6, all drawn to the same height scale. One die: six short equal bars. Two dice: a triangle of bars peaked at 3.5. Five dice: a taller, narrower hump. Ten dice: the tallest, narrowest bell. A dashed yellow line at 3.5 runs through all four. The standard deviations printed at the left are 1.71, 1.21, 0.76 and 0.54.
Read the four rows from top to bottom. One die gives six equal bars. Two dice give a triangle: a mean of 3.5 can be made six ways, a mean of 1 only one way. Five dice give a hump, and ten dice give a bell so close to normal that the naked eye cannot tell the difference.
Three changes are visible, and they are the three the book lists. The centre stays at 3.5 in every row. The spread shrinks: the standard deviations printed at the left fall from 1.71 to 0.54. And the graph gets steeper and thinner, climbing higher as it narrows, because the total area stays one.
Nothing about a die is bell-shaped, yet by ten dice the mean is. That is the central limit theorem, demonstrated on the simplest population there is.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 366 — what happens as the number of dice increases
Intuition
Why does averaging produce a hump in the middle at all? Look at the smallest case, two dice. There are 36 equally likely rolls, and the grid lists the mean of every one of them.
Figure (svg): A six by six grid: rows are the first die from 1 to 6, columns the second die from 1 to 6, and each cell shows the mean of the two faces. The anti-diagonal of cells whose mean is 3.5 is highlighted yellow, six cells. The two corner cells with means 1 and 6 are red. Cells near the yellow diagonal are shaded more strongly than cells near the corners.
The mean is constant along each diagonal running from bottom left to top right. The diagonal for 3.5 holds six cells: 1 and 6, 2 and 5, 3 and 4, and the same pairs reversed. The corners hold one cell each: the only way to average 1 is to roll two ones.
\[ P(\bar{x} = 3.5) = \frac{6}{36} = \frac{1}{6}, \qquad P(\bar{x} = 1) = \frac{1}{36} \]
So the middle value is six times as likely as the extremes, and that is the triangle you saw in the two-dice row. Nothing about the dice favours the middle; there are simply more ways to make it, because a high face and a low face cancel.
With more dice the counting gets more lopsided. With ten dice, a mean of 1 still needs every die to show a one, a single roll out of more than 60 million, while a mean near 3.5 can be made in millions of ways. The bell shape is the shape of that counting, and it is the same whatever the population, which is why the theorem does not care what the population looks like.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 366 — the means of two dice in the collaborative exercise
Worked example
Predict the centre and the spread of the ten-dice mean from one die alone, then check the prediction against the exact distribution in the picture. First find the die's own mean and standard deviation.
The mean of one die
Why: each face has probability one sixth.
\[ \mu = \frac{1 + 2 + 3 + 4 + 5 + 6}{6} = 3.5 \]
Square each distance from 3.5
Why: the distances are 2.5, 1.5 and 0.5 on each side.
\[ 2(2.5^{2} + 1.5^{2} + 0.5^{2}) = 2(8.75) = 17.5 \]
Average the squares: the variance
Why: six equally likely faces.
\[ \sigma^{2} = \frac{17.5}{6} \approx 2.917 \]
Take the square root: the standard deviation
Why: one die's spread.
\[ \sigma = \sqrt{2.917} \approx 1.708 \]
Now the theorem's prediction for the mean of ten dice: the same centre, and a spread equal to the die's standard deviation divided by the square root of ten.
Centre of the ten-dice mean
Why: unchanged by averaging.
\[ \mu_{\bar{X}} = 3.5 \]
Spread of the ten-dice mean
Why: divide by the square root of n, with n equal to 10.
\[ \sigma_{\bar{X}} = \frac{1.708}{\sqrt{10}} = \frac{1.708}{3.162} \approx 0.540 \]
Why is it the square root of ten and not ten itself? Part 2 derives it. For now, trust the arithmetic and test it.
Read the two results together. A single die lands, on average, about 1.7 away from 3.5; the mean of ten dice lands only about half a point away. Ten dice do not make the answer ten times steadier, only about three times, because the square root of ten is about 3.16.
Check against the exact distribution
Why: node computed the standard deviation of the ten-dice row by listing all 60,466,176 rolls, and it is 0.540, matching the formula. The same check holds on every row: 1.708 over the square roots of 2 and 5 gives the 1.21 and 0.76 printed there.
\[ \frac{1.708}{\sqrt{2}} \approx 1.21, \quad \frac{1.708}{\sqrt{5}} \approx 0.76, \quad \frac{1.708}{\sqrt{10}} \approx 0.54 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 366 — the dice demonstration
Concept
Now the general statement. Suppose X is a random variable with any distribution at all, known or unknown, with mean μ and standard deviation σ. Draw random samples of size n and compute each sample's mean.
\[ \bar{X} \sim N\left(\mu_{X},\ \frac{\sigma_{X}}{\sqrt{n}}\right) \]
As n increases, the sample mean tends to be normally distributed, centred at the population mean, with standard deviation equal to the population standard deviation divided by the square root of n. The book's words: the mean has the same mean as the original distribution and a variance equal to the original variance divided by the sample size.
sampling distribution of the mean — The distribution of the sample mean over all possible random samples of size n from a population. The central limit theorem says it approaches a normal distribution as n increases.
Three claims are packed in there, and each is worth separating. The centre does not move: averaging does not bias you. The spread shrinks, and by a known amount. And the shape becomes normal, whatever shape the population had. The third claim is the surprising one, and it is why the book calls this theorem one of the most powerful ideas in all of statistics.
Sampling is done with replacement, or from a population so large that it makes no difference, so the values in each sample are independent.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 365-367 — the statement of the theorem
Notation
The whole theorem fits in one line of notation. Read it piece by piece, because each piece answers a question you will be asked.
Annotate
On: \( \bar{X} \sim N\left(\mu_{X},\ \frac{\sigma_{X}}{\sqrt{n}}\right) \)
The subscripts are there to stop a common confusion. The book writes the mean and standard deviation of X with a subscript X, and those of the sample mean with a subscript X-bar. The two means are equal; the two standard deviations are not.
\[ \mu_{\bar{X}} = \mu_{X}, \qquad \sigma_{\bar{X}} = \frac{\sigma_{X}}{\sqrt{n}} \]
As with every normal distribution in this book, the second number inside N is a standard deviation, never a variance.
Try reading the line aloud for Example 7.1, with mean 90, standard deviation 15 and samples of 25: the mean of a sample of 25 is approximately normal, centred at 90, with standard error 15 over 5, which is 3. If you can say that sentence for a new problem, you have set it up.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 366-367 — the notation and the standard error of the mean
Picture it
Dice are symmetric, which makes the theorem look easy. Here is a harder test: an exponential population, the waiting-time distribution from Chapter 5, with mean 1 and standard deviation 1. It is as lopsided as distributions get: tallest at zero with a long tail to the right.
Figure (svg): On one axis from 0 to 3: a red exponential curve starting at height 1 and falling steadily, the population. An orange curve for the mean of 2 values rises from zero, peaks near 0.5 and has a long right tail. A purple curve for 5 values peaks near 0.8, more symmetric. A tall blue curve for 30 values is a narrow bell centred at 1, lying almost exactly on a dashed grey normal curve N of 1 and 0.18.
Each coloured curve is the exact distribution of the mean of n waiting times. At n equal to 2 the peak has already moved off zero, though the right tail is still long. At n equal to 5 the curve is nearly symmetric. At n equal to 30 it is a narrow bell centred at 1, lying almost on top of the dashed normal curve with mean 1 and standard deviation 1 over the square root of 30.
\[ \frac{1}{\sqrt{30}} \approx 0.183 \]
This is what the phrase any distribution means. The population does not have to be anything like a bell. It needs only a mean and a standard deviation, and a sample big enough to average away its shape.
Notice also what skew costs you: the dice looked normal by ten, but this population needs more. The further the population is from normal, the larger n must be.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 365 — if the original population is far from normal, more observations are needed
Tweak it
The slider draws the exact distribution of the mean of n waiting times from the exponential population with mean 1. Start at n equal to 1, which is the population itself, and increase n one step at a time.
Parameter explorer
Start at n = 1 and read where the curve peaks. Increase n to 5, then 15, then 40. Where is the peak now, is the curve symmetric, and roughly how wide is it?
\[ \bar{X} = \text{mean of } n = {n} \text{ waiting times} \]
Watch two things move at once. The peak slides towards 1, the population mean, and the curve narrows. At n equal to 40 almost all of the area lies between about 0.6 and 1.4, which is the mean plus or minus about 2.5 standard errors.
\[ n = 40: \quad \sigma_{\bar{X}} = \frac{1}{\sqrt{40}} \approx 0.158 \]
Something that does not move is worth noticing too: the balance point of every curve is exactly 1. The peak starts left of 1 because the curve is skewed, but the mean of the sample mean is the population mean for every n, large or small.
Concept
The theorem says tends to be normal as n increases, which raises the practical question: how big must n be before you may use the normal curve? The book gives a working rule.
The first case explains Example 7.2 later in this deck, where soccer match times are already normal and nothing has to be waited for. The second case is the one you will use most, as in Example 7.1, where the population is unknown and n is 25. The book uses it there all the same; with 25 values and no sign of strong skew, that is a reasonable judgement, but it is a judgement.
Thirty is a convention, not a law of nature. The dice, which are symmetric, looked normal by five. The exponential waiting times, heavily skewed, still showed a visible tail at five. A strongly skewed income distribution may need hundreds.
So the question to ask of any problem is: what do I know about the population's shape, and how many values are being averaged?
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 365 — the sample size should be at least 30 or the data should come from a normal distribution
Sorting
For each situation, decide whether the normal model for the sample mean is justified by what you know. Use the book's rule: a normal population, or n of at least 30.
Sort into buckets
Sort each situation by whether the sample mean may be treated as normally distributed.
The first item surprises people: four is tiny, but when the population is normal the mean is exactly normal for any n, so size is not the issue. The wedge-shaped incomes are the opposite case, badly skewed, but a thousand values is far more than enough.
The unknown population with n of 10 is the honest no. The theorem is still true, but you cannot say how close to normal the mean is yet, so the normal-curve answer would be a guess.
Concept
Every problem in this chapter involves two random variables, and naming both in words before you compute is the book's habit for good reason. X is one value from the population. The sample mean is the average of n such values.
| X | X̄ | |
|---|---|---|
| what it is | one value from the population | the mean of a sample of n values |
| centre | μ | μ, the same |
| spread | σ | σ/√n, the standard error |
| shape | whatever the population has | approximately normal for large n |
Because the sample mean has its own spread, it has its own z-score. The distance from the mean is measured in standard errors rather than in standard deviations of single values.
\[ z = \frac{\bar{x} - \mu_{X}}{\left(\frac{\sigma_{X}}{\sqrt{n}}\right)} \]
The book puts it this way: the random variable X-bar has a different z-score associated with it from that of the random variable X. The numerator is the familiar distance from the centre; only the denominator has changed. That one change is the difference between every right and wrong answer in Parts 4 and 5.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 366-367 — the z-score for the sample mean
Section
Part 2
Concept
The first of the theorem's three claims is the simplest to see, and it holds for every sample size, not just large ones. Each value in a sample is drawn from the same population, so each has mean μ. Average n of them.
The mean of a total is the total of the means
Why: each of the n values contributes μ.
\[ \text{mean of } (X_{1} + \cdots + X_{n}) = n\mu \]
Divide the total by n
Why: dividing a quantity by n divides its mean by n.
\[ \mu_{\bar{X}} = \frac{n\mu}{n} = \mu \]
So the sample mean is centred exactly on the population mean. In the book's words, the sample mean tends to target the population mean. Statisticians call such an estimate unbiased: over many samples it is right on average, neither too high nor too low.
You saw this in every picture so far. All four rows of the dice sat on 3.5, and every curve of the exponential slider balanced at 1, even when its peak was well to the left. The peak and the balance point are different things for a skewed curve; it is the balance point, the mean, that never moves.
Keep the two facts apart as you go on. The centre of the sample mean never depends on n. The spread always does, which is the next slide.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 366-367, 369 — the mean of the sample means equals the population mean
Concept
The book says the variance of the sample mean is the population variance divided by n, and the standard deviation is its square root. You can see where that comes from in four lines, using one fact about independent values: their variances add.
The total of n independent values
Why: n variances of σ squared add up.
\[ \text{Var}(X_{1} + \cdots + X_{n}) = n\sigma^{2} \]
Divide the total by n to get the mean
Why: dividing a quantity by n divides its variance by n squared.
\[ \text{Var}(\bar{X}) = \frac{n\sigma^{2}}{n^{2}} \]
Cancel one n
Why: the variance of the mean.
\[ \text{Var}(\bar{X}) = \frac{\sigma^{2}}{n} \]
Take the square root
Why: the standard deviation of the mean.
\[ \sigma_{\bar{X}} = \frac{\sigma}{\sqrt{n}} \]
So it is the variance that divides cleanly by n, and the standard deviation inherits a square root. The intuition: when you add n values, the highs and lows partly cancel, so the total's spread grows only like the square root of n, not like n. Dividing by n then leaves one over the square root of n.
You will not need this derivation on a test, but it answers the question students always ask: why not just divide by n?
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 366 — a variance equal to the original variance divided by the sample size
Concept
standard error of the mean — The standard deviation of the sample mean, σ/√n. It describes how far, on average, a sample mean will be from the population mean in repeated random samples of size n.
That definition is the book's, and it is worth seeing. Here is a simulation from a skewed population with mean 90 and standard deviation 15, the numbers of Example 7.1. The upper strip shows 70 single values; the lower strip shows 70 means of 25 values each.
Figure (svg): Two strips of dots on one axis from 70 to 170, with a dashed yellow line at the population mean 90. The upper orange strip, 70 single values from a right-skewed population, is spread from about 75 to beyond 150, bunched on the left. The lower blue strip, 70 means of 25 values each, is packed tightly between about 83 and 98.
The single values scatter with a standard deviation of 10.2, near the population's 15. The means of 25 scatter with a standard deviation of 2.9, near the standard error the theorem predicts.
\[ \frac{15}{\sqrt{25}} = \frac{15}{5} = 3 \]
So any one sample mean of 25 is typically about 3 units from the truth, while any one value is typically about 15 away. That is why a sample tells you far more about μ than one observation does, and it is why the next chapter's margins of error are built from this number.
The name is chosen for that use. It is called an error because it measures how wrong a sample mean is likely to be as an estimate of μ.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 367-368 — the standard error of the mean
Picture it
Fix the population's standard deviation at 15 and draw the standard error for every sample size from 1 to 100. The yellow dots sit at the perfect squares, where the arithmetic is clean.
Figure (svg): A curve falling steeply from 15 at n equals 1, then flattening. Dots mark n equals 1, 4, 9, 16, 25, 36, 49, 64, 81 and 100, with the values 15, 7.5, 5, 3.75, 3 and 1.5 labelled at n equals 1, 4, 9, 16, 25 and 100.
The curve is steep on the left and nearly flat on the right. The first few observations buy a great deal of precision; later ones buy very little. Going from 1 value to 25 cuts the standard error by a factor of five. Going from 25 to 100, four times as many values, only halves it.
\[ \frac{15}{\sqrt{1}} = 15, \quad \frac{15}{\sqrt{25}} = 3, \quad \frac{15}{\sqrt{100}} = 1.5 \]
This shape governs every real study. To halve the uncertainty of an average you need four times the data; to cut it to a tenth you need a hundred times the data. That is why national surveys stop at a few thousand people: past that point, extra interviews cost a great deal and gain almost nothing.
Notice the point at n equal to 1: the standard error equals σ. A sample of one is a single value, and its mean has exactly the population's spread.
Worked example
An unknown distribution has mean 90 and standard deviation 15, and samples of size 25 are drawn. Find the value that is two standard deviations above the expected value, 90, of the sample mean.
\[ \mu_{X} = 90, \quad \sigma_{X} = 15, \quad n = 25 \]
Name the variable
Why: the question is about the sample mean, so its spread is the standard error.
\[ \bar{X} \sim N\left(90, \frac{15}{\sqrt{25}}\right) \]
Compute the standard error
Why: the square root of 25 is 5.
\[ \frac{15}{\sqrt{25}} = \frac{15}{5} = 3 \]
Use value equals mean plus a number of spreads
Why: two standard deviations of the sample mean.
\[ \text{value} = 90 + (2)(3) \]
Evaluate
Why: 90 plus 6.
\[ \text{value} = 96 \]
Figure (svg): Two curves centred at 90 on an axis from 40 to 140: a low wide orange curve for one value with sigma 15, and a tall narrow blue curve for the mean of 25 with standard error 3. A yellow line at 96 reaches the narrow curve, labelled mean plus 2 SE equals 96. A pink line at 120 reaches the wide curve, labelled one value plus 2 sigma equals 120.
So a sample mean of 96 is two standard deviations of the sample mean above 90. The book says two standard deviations; strictly these are standard errors, the standard deviation of the sample mean, and the picture shows why the distinction matters.
Two standard deviations above the mean for a single value would be 90 plus 2 times 15, which is 120. The same phrase gives 96 for a mean of 25 values and 120 for one value, so always say which variable you mean.
Check with the z-score formula for the sample mean
Why: 96 must come out exactly two standard errors above 90.
\[ z = \frac{96 - 90}{3} = \frac{6}{3} = 2 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 368 — Example 7.1(b)
Error analysis
Four students computed the standard error for Example 7.1, where the population standard deviation is 15 and the sample size is 25. Only one is right. Find the fault in each of the others.
Annotate
On: \( \frac{15}{25} \qquad \frac{15}{\sqrt{15}} \qquad 15 \qquad \frac{15}{\sqrt{25}} \)
There is a quick sanity check that catches the first and third errors at once. For any sample of more than one value, the standard error must be smaller than σ, but only by a factor of the square root of n. Here that factor is 5, so the answer must be 15 divided by 5.
\[ \frac{\sigma}{n} < \frac{\sigma}{\sqrt{n}} < \sigma \quad \text{for } n > 1 \]
The book's homework states this as a range check: every standard error lies between σ over n and σ itself. An answer outside that range has gone wrong somewhere.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 367-368 — Example 7.1 and the standard error
Trap
A population has σ equal to 15 and samples of 100 are taken. A first attempt at the standard error writes:
\[ \sigma_{\bar{X}} = \frac{15}{100} \]
\[ \sigma_{\bar{X}} = 0.15 \]
Wrong. That says means of 100 vary a hundred times less than single values.
The variance divides by n, so the standard deviation divides by the square root of n.
\[ \sigma_{\bar{X}} = \frac{15}{\sqrt{100}} \]
\[ = \frac{15}{10} \]
\[ = 1.5 \]
This is Example 7.3's standard error. The wrong answer is ten times too small, and every probability built on it would claim a precision the data do not have: a sample mean of 30 would sit 27 standard errors from 34 instead of 2.67.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 369 — Example 7.3(a)
Estimation
A study with 400 participants reports a standard error that its funders think is too large. They want it halved. Estimate the sample size needed before computing anything.
Predict first
A study with n of 400 wants to halve its standard error. What sample size does it need?
Correct: 1,600
Why: The standard error is sigma over the square root of n, so halving it means doubling the square root of n. Doubling a square root means multiplying n by four, and four times 400 is 1,600.
\[ \frac{\sigma}{\sqrt{1600}} = \frac{\sigma}{40} = \frac{1}{2} \cdot \frac{\sigma}{20} = \frac{1}{2} \cdot \frac{\sigma}{\sqrt{400}} \]
The answer of 800 is the one intuition gives: double the data, halve the error. It only cuts the standard error by a factor of the square root of 2, about 1.41. The answer of 200 goes the wrong way entirely: a smaller sample gives a larger standard error.
Four times the cost for twice the precision is the price of the square root, and it is why the curve on the previous slide flattens so fast.
Ranking
Each item gives a population standard deviation and a sample size. Order them from the smallest standard error to the largest. Be careful: the biggest σ does not automatically give the biggest standard error.
Work each one out roughly before you drag anything. Square roots of perfect squares are quick: 4, 25 and 400 give 2, 5 and 20. For 30 and 38, a square root a little above 5 and a little above 6 is close enough to rank.
Put in order
Order these from smallest standard error to largest.
Why: The standard errors are 1 for sigma 2 with n of 4, about 1.46 for sigma 8 with n of 30, 1.5 for sigma 30 with n of 400, 3 for sigma 15 with n of 25, about 8.92 for sigma 55 with n of 38, and 12 for sigma 12 with a sample of one.
\[ \frac{2}{\sqrt{4}} = 1, \quad \frac{8}{\sqrt{30}} \approx 1.46, \quad \frac{30}{\sqrt{400}} = 1.5, \quad \frac{15}{\sqrt{25}} = 3, \quad \frac{55}{\sqrt{38}} \approx 8.92 \]
The largest population spread, 30, ends up in the middle because 400 values shrink it by a factor of 20. And the smallest spread of all but one, 12, finishes last because a sample of one does not shrink it at the size of a single value.
The lesson: the standard error depends on both numbers together, and the sample size acts through its square root.
Fill the middle
Yoonie reviews 16 employees each month. One review takes her about four hours, with a population standard deviation of 1.2 hours, and review times are normally distributed. Fill in the distribution of one review and of the mean of a month's 16 reviews.
Fill in the blanks
X \sim N(4,\ 1.2), \qquad \bar16 \sim N\left(4,\ \frac0.3___} }}\right) = N(4,\ ___)
Why: One review has the population spread, 1.2 hours. The month's mean averages 16 reviews, so the standard error is 1.2 divided by the square root of 16, which is 1.2 over 4, or 0.3 hours.
The centre is 4 hours in both, because averaging does not move the centre. Only the spread changes, from 1.2 hours for one review to 0.3 hours for the mean of sixteen.
Because review times are themselves normal, the mean is exactly normal, even though 16 is well under 30. That is the first case of the large-enough rule.
Practice Exercise 1 asks you to list the three numbers first: mean 4 hours, standard deviation 1.2 hours, sample size 16. Writing them down before building either distribution is a good habit.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 practice, p. 389 — Exercises 1-2
Prediction
Push the formula to its smallest case. A sample of size one is a single value drawn from the population. What does the standard-error formula give, and does that make sense?
Predict first
For a sample of size 1, what is the standard error of the mean?
Correct: σ, the population standard deviation
Why: The square root of 1 is 1, so the standard error is sigma divided by 1, which is sigma. A sample of one value is just one value, so its mean has exactly the spread of single values.
\[ n = 1: \quad \frac{\sigma}{\sqrt{1}} = \sigma \]
This is a useful sanity check on the formula: at n equal to 1 it must reduce to the population, and it does. The zero option would be right for the other extreme, a sample that is the whole population, where the mean is μ exactly and does not vary.
Keep this case in mind for Part 4: a probability about one value is the same as a probability about a mean with n equal to one, which is why the individual-value curve always looks wider.
Section
Part 3
Concept
The book is emphatic about one sentence: the variable n is the number of values that are averaged together, not the number of times the experiment is done. Here is why the distinction is easy to miss.
Figure (svg): Ten rows labelled roll 1 to roll 10, each holding five small dice, with an arrow from each row to its mean on the right. The first row is boxed in yellow and labelled one sample equals one roll of 5 dice; underneath, n equals 5, the values inside one mean. An orange bracket down the column of ten means is labelled 10 means, the number of samples, not n.
In the collaborative exercise, someone rolls five dice, ten times over. Two numbers appear: 5 and 10. Each roll averages five faces into one mean, so n is 5. Rolling ten times gives ten means, which fill the histogram but do not change how variable each mean is.
\[ \sigma_{\bar{X}} = \frac{1.708}{\sqrt{5}} \approx 0.764 \quad \text{not} \quad \frac{1.708}{\sqrt{10}} \]
Here is a way to test it. If the person rolled a hundred times instead of ten, would each individual mean of five dice become any less variable? No: each mean is still five faces averaged. More repetitions just draw the histogram more sharply. So the number of repetitions cannot belong in the standard error.
Read across a row to find n. Read down the column to count samples.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 366 — n is the number of values averaged together
Matching
Each setup mentions at least two counts. Match each to its sample size n, the number of values averaged into one mean.
Match the pairs
Why: In every case n is the number of values inside one average: five dice, 43 songs, 16 reviews, 34 cans, 100 users. The other counts, eight students, ten rolls, five albums, twelve months, 5,000 cans an hour and twelve cities, count samples, groups or the population, not values averaged.
The albums are the trickiest. The sample was chosen as five albums, but the mean is taken over songs, and there are 43 of them, so n is 43. Always ask what the average is an average of.
The factory's 5,000 cans an hour is the population's size, which does not appear in the standard error at all, provided the sample is a small fraction of it.
The cities are the mirror image of the dice. Twelve samples of 100 give twelve means, each built from 100 users, so every one of them has standard error σ over 10. Collecting more cities would give you more means to look at, not more precise ones.
Trap
A student rolls five dice ten times and asks for the standard deviation of the mean of the faces. A first attempt uses the ten rolls:
\[ n = 10 \]
\[ \frac{1.708}{\sqrt{10}} \approx 0.540 \]
Wrong. Ten is how many means were collected.
Each mean is the average of five faces, so n is five.
\[ n = 5 \]
\[ \frac{1.708}{\sqrt{5}} \]
\[ \approx 0.764 \]
The wrong answer is the spread for ten dice, a different experiment. Check against the exact picture of Part 1: the five-dice row has standard deviation 0.76, whatever number of times it is rolled.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 366 — n is not the number of times the experiment is done
Worked example
Song lengths in a collection are uniformly distributed from 2 to 3.5 minutes. Five albums are picked at random, holding 43 songs in all. Find the distribution of the average song length and its first quartile.
Name X in words
Why: one value from the population.
\[ X = \text{length of one song, minutes}, \quad X \sim U(2, 3.5) \]
Name the sample mean in words and find n
Why: the average is over songs, so n is 43, not 5.
\[ \bar{X} = \text{mean length of the 43 songs}, \quad n = 43 \]
Mean of the uniform
Why: the midpoint.
\[ \mu = \frac{2 + 3.5}{2} = 2.75 \]
Standard deviation of the uniform
Why: the Chapter 5 formula.
\[ \sigma = \frac{3.5 - 2}{\sqrt{12}} = \frac{1.5}{3.464} \approx 0.433 \]
Standard error
Why: divide by the square root of 43.
\[ \sigma_{\bar{X}} = \frac{0.433}{\sqrt{43}} = \frac{0.433}{6.557} \approx 0.0660 \]
Write the distribution
Why: 43 values from a symmetric population is plenty.
\[ \bar{X} \sim N(2.75,\ 0.0660) \]
Figure (svg): Left: a flat orange rectangle from 2 to 3.5 minutes, the uniform distribution of one song length, titled X is U of 2 and 3.5 with sigma 0.433. Right, on a much narrower axis from 2.55 to 2.95: a blue bell centred at 2.75 with the middle half, from 2.71 to 2.79, shaded; titled X-bar is N of 2.75 and 0.0660.
Look at the two panels. The population is flat, with every length from 2 to 3.5 minutes equally likely, yet the mean of 43 songs is a clear bell, on an axis less than a sixth as wide. That is the theorem on a population that has no peak at all.
For the first quartile, 25 percent of the area lies to the left: the z-score with left area 0.25 is minus 0.674, so walk that many standard errors down from 2.75. Using n of 5 instead would give a standard error of 0.194, three times too big.
First quartile
Why: k equals the mean plus z times the standard error.
\[ Q_{1} = 2.75 + (-0.674)(0.0660) \approx 2.71 \]
Check against the book and by symmetry
Why: the book's answers are N(2.75, 0.0660) and 2.71 minutes. The third quartile is the mirror image, 2.79, and the gap between them, 0.09 minutes, matches the book's interquartile range.
\[ Q_{3} = 2.75 + (0.674)(0.0660) \approx 2.79, \quad 2.79 - 2.71 \approx 0.09 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 393 — Exercise 65
Step zero
The IRS says the average time to complete Form 1040 is 10.53 hours. The distribution is unknown; assume a standard deviation of 2 hours. A random sample of 36 taxpayers is taken. The book's exercise begins by asking you to name the two random variables in words.
Discussion prompt
Before computing anything, write X in words, write the sample mean in words, and give the distribution of the sample mean. Say why the normal model is justified even though the population is unknown.
Writing the two variables in words is not busywork. It forces you to decide whether the question is about one taxpayer or about the average of 36, and that decision picks the spread: 2 hours or one third of an hour.
Keep this setup: Part 4 returns to it and asks whether an average above 12 hours would be surprising, and whether one taxpayer above 12 hours would be. The two answers are very different.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 392 — Exercise 63(a)-(c)
Section
Part 4
Concept
Once you know the sample mean is normal, with centre μ and spread σ over the square root of n, a probability about it is an ordinary normal-curve area. The only new move is computing the spread.
\[ z = \frac{\bar{x} - \mu}{\sigma / \sqrt{n}} \]
The book does the area on a TI-83 or TI-84 with normalcdf, entering the lower value, the upper value, the mean, and the standard error. Notice the fourth entry: it is σ over the square root of n, not σ. The calculator cannot tell which you meant, so it will happily give a wrong answer if you enter the population's spread.
\[ \text{normalcdf}\left(\text{lower},\ \text{upper},\ \mu,\ \frac{\sigma}{\sqrt{n}}\right) \]
Without a calculator, the z-scores and a standard normal table give the same areas; this deck shows both the z-scores and the area so you can do either.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 367 — using the calculator to find probabilities for means
Notation
This formula is the only new piece of machinery in the section, and every symbol in it is one you have met. Read it slowly once, because the denominator is where marks are lost.
Annotate
On: \( z = \frac{\bar{x} - \mu_{X}}{\sigma_{X} / \sqrt{n}} \)
Compare it with the Chapter 6 z-score. The numerator is identical. The denominator has shrunk from σ to σ over the square root of n, so the same distance from the centre gives a larger z-score for a mean than for one value.
\[ \frac{85 - 90}{15} \approx -0.33 \quad \text{one value} \qquad \frac{85 - 90}{3} \approx -1.67 \quad \text{mean of 25} \]
That larger z-score is the theorem at work. Five units below the centre is nothing unusual for a single value, but it is well out in the tail for an average of 25.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 366-367 — the z-score for the random variable X-bar
Worked example
An unknown distribution has mean 90 and standard deviation 15. Samples of size 25 are drawn randomly. Find the probability that the sample mean is between 85 and 92.
Name the variables
Why: one value, and the mean of 25.
\[ X = \text{one value}, \quad \bar{X} = \text{mean of a sample of 25} \]
Apply the theorem
Why: the standard error is 15 over 5.
\[ \bar{X} \sim N\left(90, \frac{15}{\sqrt{25}}\right) = N(90,\ 3) \]
Standardise the lower boundary
Why: 85 is 5 below the centre.
\[ z_{1} = \frac{85 - 90}{3} = \frac{-5}{3} \approx -1.67 \]
Standardise the upper boundary
Why: 92 is 2 above.
\[ z_{2} = \frac{92 - 90}{3} = \frac{2}{3} \approx 0.67 \]
Find the area between them
Why: left area at the upper z minus left area at the lower z.
\[ P(85 < \bar{x} < 92) = 0.7475 - 0.0478 = 0.6997 \]
Figure (svg): A bell curve for the sample mean with centre 90 and standard error 3, x-bar ticks from 81 to 99 in steps of 3 and z values minus 3 to 3 beneath. The area between orange lines at 85 and 92 is shaded blue and labelled 0.6997.
So the probability that the sample mean is between 85 and 92 is 0.6997: about 70 percent of samples of 25 values give a mean in that band. The book's calculator line is normalcdf of 85, 92, 90 and 15 over the square root of 25.
Before trusting the area, look at the picture. The band from 85 to 92 contains the peak and more of the right shoulder than the left, so an answer somewhat above one half, and well short of the 95 percent of two full standard errors each side, is what the shape predicts.
Notice what this example never needed: the shape of the population. You could not answer the same question for one value, because the population is unknown. The theorem is what makes the question answerable.
Check against a node integration of the normal curve
Why: the exact area is 0.69972, which rounds to the book's 0.6997; the table values above were read at z of minus 1.667 and 0.667, not at the rounded z-scores.
\[ \Phi(0.667) - \Phi(-1.667) = 0.74751 - 0.04779 = 0.69972 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 367-368 — Example 7.1(a)
Worked example
An unknown distribution has mean 45 and standard deviation 8. Samples of size 30 are drawn randomly. Find the probability that the sample mean is between 42 and 50.
Apply the theorem
Why: n of 30 meets the rule for an unknown population.
\[ \bar{X} \sim N\left(45, \frac{8}{\sqrt{30}}\right) \]
Compute the standard error
Why: the square root of 30 is about 5.477.
\[ \frac{8}{\sqrt{30}} = \frac{8}{5.477} \approx 1.461 \]
Standardise 42
Why: 3 below the centre.
\[ z_{1} = \frac{42 - 45}{1.461} \approx -2.05 \]
Standardise 50
Why: 5 above the centre.
\[ z_{2} = \frac{50 - 45}{1.461} \approx 3.42 \]
Find the area between them
Why: almost all of the upper tail is included.
\[ P(42 < \bar{x} < 50) = 0.99969 - 0.01999 \approx 0.9797 \]
So about 98 percent of samples of 30 values have a mean between 42 and 50. The band is not centred on 45: it reaches 3.42 standard errors above but only 2.05 below, so nearly all the missing 2 percent is in the left tail.
Compare the answer you would get by mistake with σ of 8 as the spread: the z-scores would be minus 0.375 and 0.625, and the probability only about 0.38. The standard error is what makes a sample mean so predictable.
Check with the calculator line
Why: normalcdf of 42, 50, 45 and 8 over the square root of 30 gives 0.9797, the same as node's integration.
\[ \text{normalcdf}\left(42,\ 50,\ 45,\ \tfrac{8}{\sqrt{30}}\right) = 0.9797 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 368 — Try It 7.1
Prediction
Soccer matches played by an over-40 league take a normally distributed time with mean 2 hours and standard deviation half an hour. Consider the band from 1.8 to 2.3 hours. Predict before computing.
Predict first
Which is more likely to land between 1.8 and 2.3 hours: the length of one match, or the mean length of 50 matches?
Correct: the mean of 50 matches, by a wide margin
Why: Both are centred at 2 hours, but the mean of 50 has a standard error of about 0.07 hours while one match has a spread of 0.5 hours. The narrow curve of the mean fits almost entirely inside the band; the wide curve of single matches spills far beyond it.
The equally-likely option confuses having the same centre with having the same distribution. Room to spread works against a single match here: more room to spread means more chance of landing outside the band.
The last option is wrong on a fact: the match times are themselves normal, so both variables are normal. The difference between them is the spread, and only the spread.
\[ \sigma_{X} = 0.5 \quad \text{versus} \quad \sigma_{\bar{X}} = \frac{0.5}{\sqrt{50}} \approx 0.0707 \]
Worked example
The length of time, in hours, it takes an over-40 group to play one soccer match is normally distributed with mean 2 hours and standard deviation 0.5 hours. A sample of 50 matches is drawn randomly. Find the probability that the sample mean is between 1.8 and 2.3 hours.
Name the variables
Why: the book defines the sample mean as the mean time to play one match; it is the mean time of the 50 matches in the sample.
\[ \bar{X} = \text{mean time of 50 matches, hours} \]
Apply the theorem
Why: the population is normal, so the mean is exactly normal.
\[ \bar{X} \sim N\left(2, \frac{0.5}{\sqrt{50}}\right) \]
Compute the standard error
Why: the square root of 50 is about 7.071.
\[ \frac{0.5}{7.071} \approx 0.0707 \]
Standardise 1.8
Why: 0.2 hours below.
\[ z_{1} = \frac{1.8 - 2}{0.0707} \approx -2.83 \]
Standardise 2.3
Why: 0.3 hours above.
\[ z_{2} = \frac{2.3 - 2}{0.0707} \approx 4.24 \]
Find the area
Why: the upper tail beyond 4.24 is negligible.
\[ P(1.8 < \bar{x} < 2.3) = 0.99999 - 0.00234 \approx 0.9977 \]
Figure (svg): Two curves centred at 2 hours on an axis from 0.4 to 3.6. The low wide orange curve for one match has its area between 1.8 and 2.3 shaded, labelled one match 0.3812. The tall narrow blue spike for the mean of 50 matches has almost all its area inside the same band, labelled mean of 50 matches 0.9977.
So the probability that the mean time of 50 matches is between 1.8 and 2.3 hours is 0.9977. For one match, the same band has probability only 0.3812: the picture shows the narrow spike of the mean fitting inside a band that catches barely a third of the wide curve.
Check with the calculator line and with node
Why: normalcdf of 1.8, 2.3, 2 and 0.5 over the square root of 50 gives 0.9977; node's integration gives 0.99765.
\[ \text{normalcdf}\left(1.8,\ 2.3,\ 2,\ \tfrac{0.5}{\sqrt{50}}\right) = 0.9977 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 368-369 — Example 7.2
Estimation
Time taken on the SAT is normally distributed with mean 2.5 hours and standard deviation 0.25 hours. A sample of 60 students is drawn. Estimate the probability that the sample mean is between two and three hours, before doing any area calculation.
Predict first
About how likely is it that the mean time of 60 students is between two and three hours?
Correct: essentially certain, 1.0000 to four places
Why: The standard error is 0.25 over the square root of 60, about 0.032 hours. Two and three hours are each half an hour from the centre, which is more than 15 standard errors. The area outside a band that wide is far too small to show in four decimal places.
\[ \frac{0.25}{\sqrt{60}} \approx 0.0323, \qquad z = \pm\frac{0.5}{0.0323} \approx \pm 15.5 \]
The answer of about 0.95 is what you get if you use σ instead of the standard error: two and three hours are exactly two standard deviations of 0.25 from 2.5, and the Empirical Rule gives 95 percent. That is the right answer for one student, not for the mean of sixty.
An answer that comes out as 1 is not a mistake. It says a mean of 60 students essentially never strays half an hour from 2.5 hours, which is the theorem's whole point.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 369 — Try It 7.2
Worked example
A study reported that the mean age of iPad users is 34 years; suppose the standard deviation is 15 years. Take a sample of 100. Find the mean and standard deviation of the sample mean, describe its distribution, and find the probability that it is more than 30 years.
(a) Centre of the sample mean
Why: the sample mean targets the population mean.
\[ \mu_{\bar{X}} = \mu = 34 \]
(a) Its standard deviation
Why: the book calls this the sample standard deviation; it is the standard error, the standard deviation of the sample means.
\[ \sigma_{\bar{X}} = \frac{15}{\sqrt{100}} = \frac{15}{10} = 1.5 \]
(b) Its shape
Why: n of 100 is large, so approximately normal whatever the shape of individual ages.
\[ \bar{X} \sim N(34,\ 1.5) \]
(c) Standardise 30
Why: 4 years below the centre.
\[ z = \frac{30 - 34}{1.5} = \frac{-4}{1.5} \approx -2.67 \]
(c) The area to the right
Why: one minus the left tail.
\[ P(\bar{x} > 30) = 1 - 0.0038 = 0.9962 \]
Figure (svg): A bell curve for the mean age of 100 iPad users, centre 34, standard error 1.5, ticks from 29.5 to 38.5 with z values minus 3 to 3. Everything to the right of an orange line at 30 years is shaded and labelled 0.9962.
So the probability that the mean age of 100 users is more than 30 years is 0.9962. The book describes 30 as the reported mean age in this particular study, one line after the reported mean was 34; read 30 simply as the cut-off being asked about.
Individual ages are almost certainly skewed right, since users can be much older than 34 but not much younger than about 10. With n of 100, that does not matter for the sample mean.
Check with the calculator line
Why: normalcdf of 30, a huge upper bound, 34 and 1.5 gives 0.9962, and node gives 0.99617.
\[ \text{normalcdf}(30,\ 10^{99},\ 34,\ 1.5) = 0.9962 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 369 — Example 7.3(a)-(c)
Worked example
Yoonie's reviews take X ~ N(4, 1.2) hours each. Find the probability that one review takes from 3.5 to 4.25 hours, then the probability that the mean of a month's 16 reviews does, and say why they differ.
One review: standardise 3.5 and 4.25
Why: use σ of 1.2 for a single value.
\[ z = \frac{3.5 - 4}{1.2} \approx -0.42, \quad z = \frac{4.25 - 4}{1.2} \approx 0.21 \]
One review: the area
Why: a narrow slice of a wide curve.
\[ P(3.5 < x < 4.25) = 0.58252 - 0.33846 \approx 0.2441 \]
Mean of 16: the standard error
Why: 1.2 over 4.
\[ \sigma_{\bar{X}} = \frac{1.2}{\sqrt{16}} = 0.3 \]
Mean of 16: standardise
Why: the same boundaries, the new spread.
\[ z = \frac{3.5 - 4}{0.3} \approx -1.67, \quad z = \frac{4.25 - 4}{0.3} \approx 0.83 \]
Mean of 16: the area
Why: the same slice, now most of a narrow curve.
\[ P(3.5 < \bar{x} < 4.25) = 0.7977 - 0.0478 = 0.7499 \]
Figure (svg): Two panels on identical axes from 0 to 8 hours. Left: a low wide orange curve for one review with the thin slice from 3.5 to 4.25 shaded, area 0.2441. Right: a tall narrow blue curve for the mean of 16 reviews with the same slice shaded, area 0.7499.
Both calculations used the same boundaries and the same centre; only the divisor changed, from 1.2 to 0.3. Dividing by a quarter of the spread made every z-score four times as large, and that alone moved the answer from a quarter to three quarters.
Exercise 5 asks what causes the difference. The two probabilities come from two different distributions: one review has a spread of 1.2 hours, while the mean of sixteen has a spread of 0.3 hours. The same interval is a sliver of the first curve and the bulk of the second.
Check against the book
Why: the book's answer to Exercise 3 is 0.2441, matching the first result; node gives 0.24405 and 0.74988.
\[ 0.2441 \quad \text{and} \quad 0.7499 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 practice, p. 389 — Exercises 3-5
Trap
Example 7.1(a): mean 90, σ of 15, n of 25; find the probability the sample mean is between 85 and 92. A first attempt standardises with σ:
\[ z = \frac{85 - 90}{15} \approx -0.33 \]
\[ z = \frac{92 - 90}{15} \approx 0.13 \]
\[ P \approx 0.184 \]
Wrong. That is the spread of one value.
The question is about a mean of 25, so the spread is the standard error, 15 over 5.
\[ z = \frac{85 - 90}{3} \approx -1.67 \]
\[ z = \frac{92 - 90}{3} \approx 0.67 \]
\[ P = 0.6997 \]
The wrong answer is also unanswerable in principle: it pretends one value is normal, but the population here is unknown. Ask first whether the question is about one value or a mean.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 367-368 — Example 7.1(a)
Worked example
Students' pocket change is exponentially distributed with mean 0.88 dollars. Pick 25 students at random. Find the probability that one student has between 80 cents and one dollar, then the probability that the average of the 25 does, and explain the difference.
The population
Why: an exponential distribution has standard deviation equal to its mean.
\[ X \sim \text{Exp, } \mu = \sigma = 0.88 \]
One student: the exponential area
Why: Chapter 5's formula, the difference of two tail areas.
\[ P(0.80 < x < 1.00) = e^{-0.80/0.88} - e^{-1.00/0.88} \]
Evaluate
Why: 0.4029 minus 0.3210.
\[ = 0.4029 - 0.3210 = 0.0819 \]
The mean of 25: the theorem
Why: the standard error is 0.88 over 5.
\[ \bar{X} \sim N\left(0.88, \frac{0.88}{\sqrt{25}}\right) = N(0.88,\ 0.176) \]
Standardise both boundaries
Why: on the sampling scale.
\[ z = \frac{0.80 - 0.88}{0.176} \approx -0.45, \quad z = \frac{1.00 - 0.88}{0.176} \approx 0.68 \]
The area
Why: a normal-curve area now.
\[ P(0.80 < \bar{x} < 1.00) = 0.7523 - 0.3247 = 0.4276 \]
Figure (svg): Two panels on identical axes from 0 to 3 dollars. Left: an orange exponential curve starting high at 0 and decaying, with the thin strip from 0.80 to 1.00 shaded, area 0.0819. Right: a blue bell centred at 0.88 with the same strip shaded, area 0.4276.
The two parts use different distributions and different methods. One student's change follows the skewed exponential curve, where 80 cents to a dollar is a thin slice. The average of 25 follows a normal curve centred at 88 cents, where the same strip is the middle of the curve. Using the normal curve for one student would be wrong twice: wrong shape and wrong spread.
Check against the book's answers
Why: the book gives 0.0819 and 0.4276 for parts (e) and (f), matching node's 0.08191 and 0.42760.
\[ 0.0819 \quad \text{and} \quad 0.4276 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 392 — Exercise 61
Comparison
Example 7.3's iPad users once more: μ of 34 years, σ of 15, samples of 100. Fill in the blanks for the sample mean, and the z-score of 30 for one user. Individual ages are probably skewed.
Comparison matrix
| one user, X | mean of 100 users, X̄ | |
|---|---|---|
| centre | 34 | 34 |
| spread | 15 | 1.5 |
| shape | probably skewed right | approximately normal |
| z-score of 30 | -0.27 | −2.67 |
| 95th percentile | cannot say without the shape | 36.5 |
The centre is shared; everything else differs. The spread falls by a factor of ten, the square root of 100. That turns 30 from an ordinary age, only about a quarter of a standard deviation below the mean, into a very unusual sample mean, two and two thirds standard errors below.
\[ \frac{30 - 34}{15} \approx -0.27 \qquad \frac{30 - 34}{1.5} \approx -2.67 \]
The last row matters most. For one user, a percentile needs the population's shape, which you do not have. For the mean of 100, the theorem supplies the shape, so the percentile is a Chapter 6 calculation: 34 plus 1.645 times 1.5.
Concept
Several questions in this section ask whether a sample mean would be surprising. The Empirical Rule from Chapter 6 answers them, applied on the sampling scale: measure the distance from μ in standard errors, and read off how often a mean lands that far out.
| distance from μ | how often a sample mean lands that far out or farther | verdict |
|---|---|---|
| within 1 SE | about 68% land inside | ordinary |
| 1 to 2 SE | about 27% land here | routine |
| beyond 2 SE | about 5%, both tails together | unusual |
| beyond 3 SE | about 0.3% | very unusual |
Apply it to the cases you have met. A mean age of 30 for 100 iPad users is 2.67 standard errors below 34: unusual, and the tail beyond it holds 0.38 percent. A mean of more than 12 hours for 36 taxpayers is 4.41 standard errors above 10.53: essentially never. A mean of 90.5 for 25 values from Example 7.1 is a sixth of a standard error from 90: completely ordinary.
\[ \frac{30 - 34}{1.5} \approx -2.67, \qquad \frac{12 - 10.53}{0.333} \approx 4.41, \qquad \frac{90.5 - 90}{3} \approx 0.17 \]
This is the reasoning that Chapter 9 formalises as a hypothesis test. Assume the population mean is what is claimed, compute how far the sample mean is in standard errors, and ask whether chance alone would often produce that. The next example is a first taste.
The yardstick only works in standard errors. The same distance measured in σ would make every sample mean look ordinary.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 369-370 — Example 7.3(c) and Try It 7.4
Worked example
Cans claim to hold 16 ounces. A sample of 34 cans has mean 16.01 ounces. If cans are filled with μ equal to 16.00 ounces, as labelled, and σ equal to 0.143 ounces, find the probability that 34 cans average more than 16.01. Do the results suggest the cans hold more than 16 ounces?
Apply the theorem
Why: n of 34 is at least 30.
\[ \bar{X} \sim N\left(16, \frac{0.143}{\sqrt{34}}\right) \]
Compute the standard error
Why: the square root of 34 is about 5.831.
\[ \frac{0.143}{5.831} \approx 0.02452 \]
Standardise 16.01
Why: a hundredth of an ounce above.
\[ z = \frac{16.01 - 16}{0.02452} \approx 0.41 \]
The area to the right
Why: one minus the left area.
\[ P(\bar{x} > 16.01) = 1 - 0.6583 = 0.3417 \]
Figure (svg): A bell curve for the mean fill of 34 cans, centred at 16 ounces with standard error 0.0245, on an axis from 15.92 to 16.08. An orange line at 16.01 is labelled z equals 0.41, and the area to its right is shaded and labelled 0.3417.
Now interpret. If the cans really average 16 ounces, a sample of 34 would average 16.01 or more about 34 percent of the time, roughly one sample in three. A sample mean of 16.01 is therefore entirely ordinary for correctly filled cans.
So no: the results do not suggest overfilling. A sample mean slightly above 16 is exactly what chance produces. Only a probability that is very small, such as the 0.0038 of Example 7.3's left tail, would make a result hard to explain by chance, and Chapter 9 turns that idea into a formal test.
Check the direction and size
Why: 16.01 is above 16, so z is positive and the right tail is less than half; 0.41 standard errors is close to the centre, so the tail should be not much below one half. node gives 0.34173.
\[ 0 < 0.41 \quad \Rightarrow \quad 0 < P = 0.3417 < 0.5 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 370 — Try It 7.4
Socratic
Example 7.1 gave a population with mean 90 and standard deviation 15 but said the distribution is unknown. It then asked about a sample mean of 25. Suppose it had asked about a single value instead.
Discussion prompt
Could you find the probability that one value from this population is between 85 and 92? What would you need to know, and why does the question about the mean of 25 not need it?
This is the practical power of the theorem. Real populations rarely come with a known shape, but they often come with a reasonable estimate of their mean and standard deviation. Averages let you work with those two numbers alone.
It is also a warning. Any question about one value from an unknown population cannot be answered with normalcdf, however tempting the numbers look. The next slide shows the book's own homework falling into exactly that.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 367 — Example 7.1: an unknown distribution
Edge cases
Return to the IRS setup: Form 1040 times with mean 10.53 hours, standard deviation 2 hours, distribution unknown, and a sample of 36 taxpayers, so the sample mean is N(10.53, 0.333).
Discussion prompt
Part (d): would you be surprised if the 36 taxpayers averaged more than 12 hours? Part (e): would you be surprised if one taxpayer took more than 12 hours? Answer each with a probability if you can, and say what you cannot compute and why.
The book's printed answer to part (e) is 0.2312, which is the normal-curve area beyond 12 with σ of 2. That treats one taxpayer's time as normal, which the exercise has just said is unknown. The conclusion, not surprising, is sensible; the number is not justified.
This is the boundary of the theorem: it speaks about averages of many values, never about one value.
Chebyshev's inequality, a result valid for every distribution, does give a rough bound for one value: at most one over k squared of values lie k or more standard deviations from the mean. For 0.735 standard deviations that bound says nothing useful, which confirms the verdict: one taxpayer at 12 hours is not surprising.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 392 — Exercise 63(d)(e)
Real world
You are developing a strategy game aimed at men from their late twenties to late thirties. Industry research says the average strategy player is 28 years old with a standard deviation of 4.8 years. You take a random sample of 100 gamers.
Discussion prompt
Find the probability that the mean age of 100 randomly chosen strategy gamers is between 29 and 35. What does this tell you about your target market, and what does it not tell you?
This is how the standard error is used in market research and polling: a sample mean is compared with a claimed population mean, measured in standard errors. A mean two or more standard errors away is a signal worth acting on.
The limit of the conclusion is the same limit as on the last slide. A statement about averages is not a statement about individuals.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 370 — Try It 7.3
Section
Part 5
Concept
A percentile turns the question round. Instead of a boundary in and a probability out, you are given an area and asked for the boundary. It is Chapter 6's x equals μ plus zσ, with the standard error in place of σ.
\[ k = \mu + z \cdot \frac{\sigma}{\sqrt{n}} \]
Two moves. First find the z-score whose left area is the percentile you want, from a table or the calculator. Then walk that many standard errors from the mean. For the 95th percentile, the z-score is 1.645; for the 90th, 1.282; for the 80th, 0.842; for the first quartile, minus 0.674.
| percentile | left area | z |
|---|---|---|
| 25th (first quartile) | 0.25 | −0.674 |
| 50th (median) | 0.50 | 0 |
| 80th | 0.80 | 0.842 |
| 90th | 0.90 | 1.282 |
| 95th | 0.95 | 1.645 |
The book uses invNorm with the area, the mean and the standard error, and again the fourth entry is the standard error. The median of the sample mean is μ itself, because the normal curve is symmetric, a fact Homework 64 asks for directly.
\[ \text{invNorm}\left(\text{area},\ \mu,\ \frac{\sigma}{\sqrt{n}}\right) \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 369 — using the calculator to find percentiles for means
Worked example
Back to the iPad users: μ of 34 years, σ of 15, samples of 100, so the sample mean is N(34, 1.5). Find the 95th percentile of the sample mean age, to one decimal place.
Name the target
Why: the sample mean age with 95 percent of sample means below it.
\[ P(\bar{X} < k) = 0.95 \]
Find the z-score
Why: the left area 0.95 gives z of 1.645.
\[ z_{0.95} = 1.645 \]
Walk that many standard errors up from 34
Why: 1.645 times 1.5.
\[ k = 34 + (1.645)(1.5) \]
Multiply
Why: the distance above the centre in years.
\[ k = 34 + 2.467 \]
Add and round to one decimal place
Why: the book's 36.5.
\[ k = 36.467 \approx 36.5 \text{ years} \]
Figure (svg): The bell curve for the mean age of 100 iPad users, centre 34 and standard error 1.5. The area to the left of a yellow line at 36.47 is shaded and labelled 0.95; the small tail to its right is labelled 0.05.
Interpret it about samples: 95 percent of all random samples of 100 iPad users have a mean age below 36.5 years. It is not a statement about individual users. The 95th percentile of individual ages would need their distribution; if they were normal it would be 34 plus 1.645 times 15, about 58.7, a very different number.
The picture shows why a percentile is the reverse of Part 4: the area is fixed at 0.95 and the line k is the unknown.
Check by standardising the answer
Why: 36.467 should be 1.645 standard errors above 34, with left area 0.95.
\[ \frac{36.467 - 34}{1.5} = 1.645, \quad \Phi(1.645) = 0.9500 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 369 — Example 7.3(d)
Picture it
Fix the population at mean 34 and standard deviation 15, and compute the 95th percentile of the sample mean for four sample sizes: 1, 4, 25 and 100. The case n equal to 1 is a single user, drawn here as if ages were normal.
Figure (svg): Four horizontal bars starting at a dashed line at 34 years. For one person the bar reaches 58.67; for the mean of 4 it reaches 46.34; for the mean of 25, 38.93; for the mean of 100, 36.47.
\[ 34 + 1.645\left(\frac{15}{\sqrt{n}}\right): \quad 58.67,\ 46.34,\ 38.93,\ 36.47 \]
Every bar is 1.645 standard errors long, but the standard error shrinks as n grows, so the percentile slides in towards 34. Quadrupling n from 1 to 4 halves the bar; quadrupling again from 25 to 100 halves it again.
The practical reading: with bigger samples, even an extreme sample mean is close to the population mean. A statement like 95 percent of sample means are within a certain distance of μ becomes a tighter statement as n grows, and that is exactly what a confidence interval in Chapter 8 will say, turned round to estimate μ.
Notice the pattern matches the diminishing returns of Part 2: the bar lengths are 24.7, 12.3, 4.9 and 2.5, the same square-root shrinkage.
Worked example
The mean app engagement by an iPad user is 8.2 minutes, with standard deviation 1 minute. Take a sample of 60. Find the mean and standard error of the sample mean, its 90th percentile with an interpretation, and the probability it is between 8 and 8.5 minutes.
(a)(b) Centre and standard error
Why: the square root of 60 is about 7.746.
\[ \mu_{\bar{X}} = 8.2, \quad \sigma_{\bar{X}} = \frac{1}{\sqrt{60}} \approx 0.1291 \]
(c) The z-score for the 90th percentile
Why: left area 0.90.
\[ z_{0.90} = 1.282 \]
(c) Walk up from 8.2
Why: 1.282 standard errors.
\[ k = 8.2 + (1.282)(0.1291) = 8.2 + 0.1654 \approx 8.37 \]
Interpret it: 90 percent of all samples of 60 iPad users have a mean engagement time below 8.37 minutes. The book writes that 90 percent of the average app engagement time for iPad users is less than 8.37 minutes, which blurs samples into users; the percentile describes sample means, and says nothing about 90 percent of users.
(d) Standardise 8 and 8.5
Why: on the sampling scale.
\[ z = \frac{8 - 8.2}{0.1291} \approx -1.55, \quad z = \frac{8.5 - 8.2}{0.1291} \approx 2.32 \]
(d) The area
Why: left area at 2.32 minus left area at minus 1.55.
\[ P(8 < \bar{x} < 8.5) = 0.98993 - 0.06067 \approx 0.9293 \]
Figure (svg): A bell curve for the mean engagement time of 60 iPad users, centre 8.2 minutes, standard error 0.129, on an axis from 7.75 to 8.65. The area between orange lines at 8 and 8.5 is shaded and labelled 0.9293. A yellow line at 8.37 is labelled 90th percentile.
So about 93 percent of samples of 60 users have a mean engagement between 8 and 8.5 minutes. The figure shows both answers: the yellow percentile line and the shaded band.
Check against the book and node
Why: the book gives 8.37 and 0.9293; node gives 8.3654 and 0.92926. The 90th percentile must also sit inside the band, above the centre, and it does.
\[ 8.2 < 8.37 < 8.5 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 370 — Example 7.4
Discrimination
Example 7.4 found that the 90th percentile of the mean engagement time of 60 users is 8.37 minutes. Sort each sentence by whether it is a correct interpretation.
Sort into buckets
Sort each sentence as a correct or incorrect interpretation of the 90th percentile 8.37 minutes.
The population mean is 8.2: it is not random, and nothing about it has a 90 percent chance. The randomness belongs to the sample, and the percentile describes where sample means fall.
The individual-user sentence is the most common slip, and it is wrong by a lot. For single users the spread is 1 minute, not 0.129, so their 90th percentile would be about 9.48 minutes, not 8.37.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, p. 370 — Example 7.4(c)
Worked example
A class of world-class runners finish a marathon in an average of 145 minutes with a standard deviation of 14 minutes. Consider 49 races. Give the distribution of the average time, the probability it is between 142 and 146 minutes, its 80th percentile and its median.
The distribution
Why: 14 over the square root of 49 is 2.
\[ \bar{X} \sim N\left(145, \frac{14}{\sqrt{49}}\right) = N(145,\ 2) \]
Standardise 142 and 146
Why: 3 below and 1 above the centre.
\[ z = \frac{142 - 145}{2} = -1.5, \quad z = \frac{146 - 145}{2} = 0.5 \]
The area between
Why: left areas at 0.5 and at minus 1.5.
\[ P(142 < \bar{x} < 146) = 0.6915 - 0.0668 = 0.6247 \]
The 80th percentile
Why: z of 0.842 standard errors above 145.
\[ k = 145 + (0.842)(2) = 145 + 1.683 \approx 146.68 \]
The median
Why: a normal curve is symmetric, so its median is its centre.
\[ \text{median} = 145 \]
Figure (svg): A bell curve for the mean time over 49 marathons, centre 145 minutes, standard error 2, on an axis from 138 to 152. The area between 142 and 146 is shaded and labelled 0.6247, and a yellow line at 146.68 is labelled 80th percentile.
So about 62 percent of 49-race averages fall between 142 and 146 minutes, and 80 percent fall below 146.68 minutes. The interval is lopsided, three below the centre and one above, which is why it catches less than the symmetric band from 143 to 147, which would hold about 68 percent.
All four answers used the standard error of 2, never σ of 14: every question in the exercise is about the average of the 49 races.
Check the percentile against the probability
Why: 146.68 is above 146 and 146 has left area 0.6915, so the 80th percentile must lie above 146, and standardising it gives back 0.842.
\[ \frac{146.68 - 145}{2} = 0.84, \quad \Phi(0.842) = 0.80 \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, pp. 392-393 — Exercise 64
Fill the middle
Yoonie's 16 monthly reviews give a sample mean that is N(4, 0.3) in hours. Find the 95th percentile of the mean time for one month's reviews. Fill in the z-score and the answer to two decimal places.
Fill in the blanks
k = 4 + (1.645)(0.3) = 4 + 0.4935 \approx 4.49
Why: The 95th percentile has left area 0.95, which is 1.645 standard deviations above the centre on a normal curve. Here the spread is the standard error, 0.3 hours, so walk 1.645 times 0.3, about 0.49 hours, above 4.
Interpret it: in 95 percent of months, the mean time for Yoonie's 16 reviews is below about 4.49 hours. Only about one month in twenty will average longer than that.
If you used σ of 1.2 you would get 5.97 hours, which is the 95th percentile of one review. A single review over six hours is not unusual one time in twenty; a month averaging that long would be astonishing.
\[ 4 + (1.645)(1.2) = 5.97 \quad \text{is for one review, not the mean} \]
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 practice, p. 389 — Exercise 6
Counterexample
A tempting reading of the standard error: a sample of 100 always gives a mean closer to μ than a sample of 25. The theorem says the larger sample's standard error is half as big. Test the claim.
Discussion prompt
With μ of 90 and σ of 15, the mean of 25 values has standard error 3 and the mean of 100 has standard error 1.5. Is it possible for a particular sample of 100 to land farther from 90 than a particular sample of 25? What does the standard error actually promise?
This distinction runs through the rest of the course. Statistics makes promises about procedures, what happens over repeated samples, and never guarantees a particular result.
A good way to hold it: the sampling distribution of the larger sample is narrower, so a far-off mean is less likely, not impossible.
You can even put a number on the unlucky case. A mean of 100 values lands more than 2 units from 90 with probability about 0.18, and a mean of 25 lands within half a unit of 90 with probability about 0.13; neither event is rare, so the reversal happens often enough to see.
Estimation
In 1940 the average U.S. farm was 174 acres; say the standard deviation was 55 acres. A random survey of 38 farmers is taken. Estimate the interquartile range of the sample mean: the band holding its middle half.
Predict first
Which band holds the middle half of the mean farm size of 38 farms?
Correct: about 168 to 180 acres
Why: The standard error is 55 over the square root of 38, about 8.9 acres. The quartiles of a normal curve are 0.674 spreads either side of the centre, so the band runs from about 174 minus 6 to 174 plus 6.
\[ \frac{55}{\sqrt{38}} \approx 8.92, \quad 174 \pm (0.674)(8.92) = 174 \pm 6.02 \]
\[ Q_{1} \approx 167.98, \qquad Q_{3} \approx 180.02 \]
The band from 137 to 211 is the middle half for single farms, if they were normal: it uses σ of 55. The band from 156 to 192 steps two standard errors either side, which is about 95 percent, not 50. And 172 to 176 divides by 38 instead of its square root.
Each wrong band is a mistake from earlier in the deck: the wrong spread, the wrong percentile, or dividing by n.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 393 — Exercise 66
Section
Part 6
Pattern
Every problem in this section ran through the same short procedure. The first step decides everything after it.
Figure (svg): A flow diagram. First box: is the question about one value or a sample mean? An arrow to: find mu, sigma and n, and write X-bar is N of mu and sigma over root n. Then branch: a probability, via z equals x-bar minus mu over the standard error; or a percentile, via k equals mu plus z times the standard error. Both lead to: sketch the curve, shade, and say it in context about samples of size n.
\[ z = \frac{\bar{x} - \mu}{\sigma / \sqrt{n}} \quad \Longleftrightarrow \quad k = \mu + z \cdot \frac{\sigma}{\sqrt{n}} \]
Run the procedure once on Example 7.1 to see it work. A mean of 25 values, so n is 25; the population is unknown but n is close to 30 with no sign of skew; the distribution is N(90, 3); the boundaries 85 and 92 become z-scores of minus 1.67 and 0.67; the area is 0.6997; and the sentence is that about 70 percent of samples of 25 have a mean between 85 and 92.
Every error in this deck happened at the first or last step: using σ for a mean, taking n as the number of samples, or interpreting a statement about sample means as a statement about individuals. The arithmetic between is the same as Chapter 6.
Check
Homework Exercise 71. Choose the distribution first; the arithmetic is one step.
Check your understanding
The cost of unleaded gasoline once followed an unknown distribution with mean 4.59 dollars and standard deviation 0.10 dollars. Sixteen stations are chosen at random. Which distribution should be used for the average cost at the 16 stations?
Answer: A
Why: The centre is the population mean, 4.59 dollars. The spread is the standard error, 0.10 divided by the square root of 16, which is 0.10 over 4, or 0.025 dollars.
Each distractor is one of this deck's traps: the population spread for a mean, dividing by n, and a centre that moves. The centre never moves.
One caution the exercise leaves implicit: with an unknown population and only 16 stations, the normal model rests on the population not being badly skewed. Gas prices across a region usually are not, so it is a reasonable choice, but it is a choice.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 393 — Exercise 71
Check
Homework Exercise 67, part (c). Read the claim carefully: it is about the spread of the sample mean.
Check your understanding
True or false: when the sample size is large, the standard deviation of the sample mean is approximately the same as the standard deviation of X. Which verdict and reason is right?
Answer: A
Why: The standard deviation of the sample mean is the standard error, sigma over the square root of n. For large n that is much smaller than sigma: with n of 100 it is one tenth of it. So the statement is false.
Parts (a) and (b) of the same exercise are both true: the mean of the sample mean equals the mean of X, and for large n the sample mean is approximately normal. Only the spread changes, and part (c) is the one that tests whether you know by how much.
The last distractor gets the verdict right for the wrong reason, which in a written answer loses the mark.
A quick way to justify the verdict with numbers: take Example 7.3, where σ is 15 and n is 100. The standard deviation of the sample mean is 1.5, a tenth of 15, which is nowhere near approximately the same.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 393 — Exercise 67
Check
Homework Exercise 69. The population is badly skewed, and the question is about the mean of a large sample.
Check your understanding
Incomes in a country are wedge-shaped, with mean 2,000 dollars and standard deviation 8,000 dollars. A random sample of 1,000 residents is surveyed. Why is the sample average more likely to fall between 2,000 and 2,100 dollars than between 2,100 and 2,200?
Answer: A
Why: With 1,000 residents the central limit theorem makes the sample mean approximately normal, centred at 2,000 with standard error 8,000 over the square root of 1,000, about 253 dollars. A normal curve is highest at its centre, so the band starting at the centre holds more area, 0.154 against 0.132.
\[ \frac{8000}{\sqrt{1000}} \approx 252.98, \quad 0.1537 \text{ versus } 0.1317 \]
The trap in this question is the word wedge-shaped. It describes single incomes, and it would be the right reason if the question asked about one resident. With a thousand residents averaged, the shape that matters is the sampling distribution's, and that is a bell centred at 2,000.
Part (d) of the same exercise asks how the standard deviation can exceed the mean. A wedge of incomes has many near zero and a long tail of a few very large ones; that tail makes the spread huge, even though no income is negative.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 393 — Exercise 69
Explain it to yourself
Two ideas carry the whole section: the sample size shrinks the spread, and it does so through a square root. Explaining both in your own words is the best test of whether they have landed.
Discussion prompt
Using the dice, explain in a few sentences what n counts, why the mean of ten dice varies less than one die, and why doubling the number of dice does not halve the spread.
A strong answer separates the two numbers in the experiment, the dice in a roll and the number of rolls, and says that only the first is n. It also names variance as the quantity that divides by n.
If your explanation mentioned that the centre stays at 3.5 throughout, that is the third piece: averaging changes the spread and the shape, never the centre.
Exit ticket
Homework Exercise 62. Fly balls hit to the outfield travel a normally distributed distance with mean 250 feet and standard deviation 50 feet. A random sample of 49 fly balls is taken.
Discussion prompt
Give the distribution of the average distance of the 49 fly balls. Find the probability that they average less than 240 feet, and the 80th percentile of the average, and interpret the percentile in a sentence.
The marks are in three places: the standard error, 50 over 7, rather than 50; the left tail at z of minus 1.4; and a percentile sentence that talks about samples of 49, not individual fly balls.
The population is normal here, so the mean is exactly normal and the size of n does not need justifying. If the same question came with an unknown population, 49 would still be enough.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1 homework, p. 392 — Exercise 62
Recap
A sample mean is a random variable with its own distribution. The central limit theorem says that distribution is approximately normal for large n, whatever the population's shape, centred at the population mean, with spread σ divided by the square root of n.
| idea | formula | read it as |
|---|---|---|
| the theorem | X̄ ~ N(μ, σ/√n) | means of n values: same centre, smaller spread, normal shape |
| standard error | σ/√n | typical distance of a sample mean from μ; four times n halves it |
| n | values averaged | not the number of samples, groups or repetitions |
| large enough | n ≥ 30, or a normal population | more if the population is strongly skewed |
| probability | z = (x̄ − μ)/(σ/√n) | Chapter 6 with the standard error as the spread |
| percentile | k = μ + z·σ/√n | so many percent of samples of size n have a mean below k |
The two decisions that matter come before and after the arithmetic: is the question about one value or a mean, and does the final sentence talk about samples of size n rather than individuals? Get those right and the rest is the normal curve you already know.
The Introductory Business Statistics text covers the same theorem in its own §7.1, with business contexts, if you want more practice. Section 7.2 applies the same theorem to sums instead of means.
OpenStax Introductory Statistics 2e, §7.1 The Central Limit Theorem for Sample Means (Averages) §7.1, pp. 365-370 — the whole section
OpenStax Introductory Business Statistics 2e, §7.1 The Central Limit Theorem for Sample Means §7.1 The Central Limit Theorem for Sample Means — the same ideas in business contexts
Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.