7.2 The Central Limit Theorem for Sums

The same theorem stated for totals rather than averages. If random samples of size n are drawn from any population with mean mu and standard deviation sigma, then as n increases the sum of the sample tends to be normally distributed, with a mean equal to the population mean multiplied by the sample size and a standard deviation equal to the population standard deviation multiplied by the square root of the sample size. The asymmetry between those two is the section's whole content: the centre of a total grows in proportion to n while its spread grows only in proportion to the square root of n, so a total becomes proportionally more predictable as more values are added. Probabilities and percentiles for a sum are then ordinary normal questions asked on that distribution, and because a sum is simply a mean multiplied by n, either form can be converted into the other as a check.

Subject: Statistics · 67 slides · symbolic lesson

Open the interactive version of this deck

What this lesson covers

The lesson, slide by slide

1. The Central Limit Theorem for Sums

Title

Statistics · Chapter 7 — The Central Limit Theorem

Totals of many values pile up into a bell whose centre grows like n and whose spread grows only like the square root of n

2. What this lesson gives you

Objectives

Section 7.1 told you how the average of a random sample behaves. This section asks the same question of the total: add up n values drawn from any population, and what does the sum look like? The answer is another normal distribution, and working with it uses the normal-curve skills you already have.

Keep one contrast in mind throughout. For an average, the square root of n sits underneath σ and shrinks the spread. For a total, it multiplies σ and grows the spread, but more slowly than the total itself grows. Most mistakes in this section come from putting that square root in the wrong place, so every example on these slides says out loud where it goes and why.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-373 — the section this deck covers

3. Before anything new: the spread of an average

Warm-up

Section 7.1 gave you the central limit theorem for means. Recall it on the numbers this deck starts from, because the theorem for sums is the same result wearing different clothes.

Discussion prompt

A population has mean 90 and standard deviation 15. You draw a random sample of 80 values. What are the mean and the standard deviation of the sample mean? Then: the total of the 80 values is 80 times their mean. Guess the mean of the total.

The centre of the total was easy: multiply by 80. Hold on to the question the warm-up did not ask, the spread of the total. If the total is 80 times the mean, is its spread 80 times 1.677? Part 2 shows that it is, and that 80 times 15 over the square root of 80 is the square root of 80 times 15, about 134.16.

Also recall the z-score from Chapter 6: a value minus the centre, divided by the spread. Every probability in this section is found by computing exactly that, with the centre and spread of the total in place of those of one value.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.1, p. 369 — the sample mean's standard deviation, σ over root n

4. The distribution of a sum

Section

Part 1

5. One sample, one total

Concept

The book's own picture of the theorem is rolling ten dice and adding them up. One roll of ten dice is one sample of size ten, and it produces one number: its sum. Here are four such samples, and then 150 of them.

Figure (svg): Four rows of ten dice faces, the first reading 2, 1, 2, 1, 1, 2, 5, 6, 5, 2 with sum 27. Below, a dot plot of 150 sums of ten dice on an axis from 15 to 55, piled highest near 35 and thinning out on both sides; the four sample sums shown above are the yellow dots.

One sample of ten dice is one value of the random variable ΣX. Repeat the sampling and the sums pile up into a hump centred near 35.

Look at the difference between the rows and the dot plot. Each row is ten values of X, the result of one die. Each yellow dot below is one value of a different random variable: the sum of a whole sample.

\[ \Sigma X = X_1 + X_2 + \cdots + X_{10} \]

Why the pile in the middle? A total near 35 can be made in a huge number of ways, with some high dice and some low ones. A total near 10 or 60 can be made in essentially one way. Chance has many more routes to the middle than to the edges.

The sums are not spread evenly from 10 to 60. They cluster near 35, with an average of 34.6 over these 150 samples, and they thin out quickly on both sides. Nobody in these 150 samples came close to 10 or to 60, which would need all ten dice to show one or all ten to show six.

ΣX — The random variable whose value is the sum, or total, of the n values in one random sample. A particular observed sum is written Σx.

So the question of this section is about that dot plot: what distribution do the sums follow, what is its centre, and how wide is it? The answer depends on only three numbers, the population's mean, its standard deviation and the sample size.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 370 — repeatedly rolling ten dice and calculating the sum of each sample

6. From a flat die to a bell-shaped total

Picture it

A single fair die is about as far from a bell as a distribution can be: six equally likely values, flat. Watch what adding does to that shape. These are exact probabilities, computed by listing every outcome, not simulations.

Figure (svg): Four bar charts of exact probabilities. One die: six equal bars, flat. Sum of two dice: a triangle peaking at 7. Sum of three dice: a rounded hump peaking at 10 and 11. Sum of ten dice: a smooth bell from 10 to 60 peaking at 35, with a yellow normal curve lying on the bar tops. Under each: the mean 3.5, 7, 10.5, 35 and the standard deviation 1.71, 2.42, 2.96, 5.40.

The population is flat, nothing like a bell. Add values and the shape rounds off fast: by ten dice the yellow normal curve and the exact bars are hard to tell apart.

With one die the bars are flat. With two, the sum is a triangle peaking at 7, because there are six ways to make 7 and only one way to make 2 or 12. With three the triangle rounds into a hump, and by ten the bars lie almost exactly under a normal curve.

Read the numbers under each panel as well as the shapes. The centre goes 3.5, 7, 10.5, 35: it is multiplied by the number of dice. The spread goes 1.71, 2.42, 2.96, 5.40: it grows, but much more slowly. Doubling the dice from one to two multiplies the spread by about 1.41, the square root of 2, not by 2.

Those two patterns, the shape turning normal and the two parameters growing at different rates, are the entire theorem. The next slide writes them down.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 370 — the sums tend to follow a normal distribution

7. The theorem for sums, stated

Concept

Let X be one value from a population with any distribution, known or unknown, with mean μX and standard deviation σX. Draw random samples of size n and add up each one.

\[ \Sigma X \sim N\big((n)(\mu_X),\ (\sqrt{n})(\sigma_X)\big) \]

In words: as n increases, the sums tend to be normally distributed. The normal curve they follow has a mean equal to the original mean multiplied by the sample size, and a standard deviation equal to the original standard deviation multiplied by the square root of the sample size.

\[ \mu_{\Sigma X} = n\mu_X, \qquad \sigma_{\Sigma X} = \sqrt{n}\,\sigma_X \]

The book's sentence here contains a slip worth knowing about. The paragraph stating the theorem for sums says that as sample sizes increase, the distribution of means more closely follows the normal distribution. In that paragraph it means the distribution of sums; means were the business of Section 7.1.

Notice what the theorem does not ask for. You never need the shape of the population, only its mean and standard deviation. That is why the dice, a flat population, and the ages of iPad users, whose shape nobody gives you, are handled by exactly the same formula.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-371 — statement of the central limit theorem for sums

8. Decoding ΣX ~ N((n)(μX), (√n)(σX))

Notation

The book writes the theorem in one dense line. Take it apart piece by piece, because every piece is a place to go wrong.

Annotate

On: \( \Sigma X \sim N\big((n)(\mu_X),\ (\sqrt{n})(\sigma_X)\big) \)

  • The random variable: the sum of the n values in one random sample. Not one value, and not the mean.
  • Is approximately normally distributed, and more nearly so as n grows.
  • The centre of the sums: the population mean multiplied by the sample size.
  • The spread of the sums: the population standard deviation multiplied by the square root of the sample size, never by n itself.
  • Parameters of the population of single values, which you are told; the formula builds the sum's parameters from them.

The subscripts are how you keep track of which random variable a parameter belongs to. μX and σX describe one value; μ with a subscript ΣX and σ with a subscript ΣX describe the total. A slide that mixes them up has usually put a population spread where a sum's spread belongs.

\[ \mu_{\Sigma X} = (n)(\mu_X) \qquad \sigma_{\Sigma X} = (\sqrt{n})(\sigma_X) \]

Say the pair aloud the way you will use it: the centre scales by n, the spread by the square root of n.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-371 — the notation ΣX ~ N((n)(μX), (√n)(σX))

9. When the theorem may be used

Concept

Like its partner for means, the theorem for sums comes with conditions. They are short, and each one does real work.

Two of the three numbers you need come from the population, not from the sample: its mean μX and standard deviation σX. The third, n, is how many values go into each total.

\[ \text{inputs: } \mu_X, \ \sigma_X, \ n \quad \Longrightarrow \quad \mu_{\Sigma X} = n\mu_X, \ \sigma_{\Sigma X} = \sqrt{n}\,\sigma_X \]

The formula for the centre and the formula for the spread are exact for any n, even n equal to 2, as the dice showed. Only the word normal is an approximation that needs a large sample. So when a question gives a small n from a skewed population, you may still quote the mean and standard deviation of the total, but you may not use the normal curve for its probabilities.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 370 — if you draw random samples of size n, then as n increases

10. Sixty waiting times from a skewed population

Prediction

Waiting times at a help desk are strongly right-skewed: most calls are answered in a minute or two, a few take twenty. Their mean is 3 minutes and their standard deviation 3 minutes. A supervisor adds up the waits of 60 randomly chosen calls. Before you see anything, predict the shape of that total.

Predict first

Across many random samples of 60 calls, what shape does the distribution of the total waiting time take?

  • approximately normal, centred at 180 minutes
  • right-skewed like the single waits, centred at 180 minutes
  • approximately normal, centred at 3 minutes
  • flat, since sixty skewed values average out to no shape at all

Correct: approximately normal, centred at 180 minutes

Why: The central limit theorem for sums says the total of a large random sample is approximately normal whatever the population's shape. Its centre is n times the population mean, 60 times 3, which is 180 minutes.

\[ \mu_{\Sigma X} = (60)(3) = 180, \qquad \sigma_{\Sigma X} = \sqrt{60}\,(3) \approx 23.24 \]

The option that keeps the skew is the natural guess, and it is the one the theorem overturns: the long right tail of single calls is diluted when sixty of them are added, because a very long call is rarely joined by many others. With 60 values from a strongly skewed population, the total is close to normal, though a little right-skew survives.

The flat option has an appealing logic, that sixty skewed values should average out into nothing, but averaging out produces a peak, not a plateau: totals near the centre can be reached in many more ways than totals far from it, exactly as with the dice.

The option centred at 3 minutes describes one call, or the average of the sixty, not their total. Check the units of the centre every time: a total of 60 waits must be in the hundreds of minutes.

11. Ten dice: the theorem checked against the exact answer

Worked example

The dice make a perfect test of the theorem, because the exact distribution of the sum of ten dice can be computed by listing outcomes. Find the theorem's centre and spread, then compare.

\[ X = \text{one die}: \quad \mu_X = 3.5, \quad \sigma_X = \sqrt{\tfrac{35}{12}} \approx 1.708, \quad n = 10 \]

Population mean

Why: the average of 1 to 6.

\[ \mu_X = \frac{1+2+3+4+5+6}{6} = \frac{21}{6} = 3.5 \]

Population variance

Why: average squared distance from 3.5.

\[ \sigma_X^2 = \frac{2.5^2 + 1.5^2 + 0.5^2 + 0.5^2 + 1.5^2 + 2.5^2}{6} = \frac{17.5}{6} = \frac{35}{12} \]

Population standard deviation

Why: take the square root.

\[ \sigma_X = \sqrt{2.9167} \approx 1.708 \]

Now the theorem. The centre of the total multiplies by n, and the spread multiplies by the square root of n.

Centre of the sum

Why: ten dice, each averaging 3.5.

\[ \mu_{\Sigma X} = (10)(3.5) = 35 \]

Spread of the sum

Why: root 10 is about 3.162.

\[ \sigma_{\Sigma X} = (\sqrt{10})(1.708) \approx 5.40 \]

Write the distribution

Why: by the central limit theorem for sums.

\[ \Sigma X \sim N(35,\ 5.40) \]

Figure (svg): Exact probability bars for the sum of ten dice from 10 to 60, peaking at 35, with the yellow curve N(35, 5.40) lying on top of them. A dashed orange line marks 35; pink arrows reach one standard deviation, 5.40, either side of it.

The formula's centre and spread are not approximations: the exact bars have mean 35 and standard deviation 5.40 to every decimal. Only the shape is approximate, and here it is very good.

The yellow curve sits on the exact bars almost perfectly. A sample of ten is small, but the die is symmetric, so there is no skew for the sum to lose.

Read the pink arrows as a ruler. About two thirds of all totals of ten dice land within 5.40 of 35, between about 29.6 and 40.4; the exact share from the bars for totals 30 to 40 is 0.687. A total of 45 or more, nearly two standard deviations up, happens in only about 4 percent of rolls. You can now say how surprising any total of ten dice is without rolling a single one.

Check against the exact distribution

Why: Computing the mean and standard deviation straight from all 60 million outcomes, by exact convolution in node, gives 35 and 5.4006: the theorem's centre and spread agree to every decimal shown.

\[ \text{exact: } \mu = 35, \quad \sigma = 5.4006 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 370 — ten dice

12. Example 7.5: the distribution of a sum of 80 values

Worked example

An unknown distribution has a mean of 90 and a standard deviation of 15. A sample of size 80 is drawn randomly from the population. Before any probability can be found, write the distribution of the sum.

\[ \mu_X = 90, \quad \sigma_X = 15, \quad n = 80 \]

Name the random variable

Why: the question is about the total of 80 values.

\[ \Sigma X = \text{the sum of 80 values} \]

Centre: multiply the mean by n

Why: 80 values each averaging 90.

\[ \mu_{\Sigma X} = (80)(90) = 7200 \]

Square root of n

Why: root 80, to four places.

\[ \sqrt{80} \approx 8.9443 \]

Spread: multiply σ by root n

Why: not by 80.

\[ \sigma_{\Sigma X} = (8.9443)(15) \approx 134.16 \]

Write the distribution

Why: the population is unknown, but n is large.

\[ \Sigma X \sim N(7200,\ 134.16) \]

Notice what you were never told: the shape of the population. The book calls it simply an unknown distribution. That is not a gap in the question; it is the point of the theorem. With 80 values, the total is close to normal whatever the population looks like, so the mean and standard deviation are all you need.

Pause on the size of that spread. The total is centred at 7,200 and its standard deviation is 134.16, under 2 percent of the centre. A single value has a standard deviation of 15 on a mean of 90, about 17 percent. Adding 80 values has made the total far more predictable, relative to its size, than any one value.

Check through the mean

Why: the total is 80 times the mean, and the mean's spread is 15 over root 80; multiplying by 80 must give the same spread.

\[ 80 \cdot \frac{15}{\sqrt{80}} = \sqrt{80}\cdot 15 \approx 134.16 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 371 — Example 7.5, the distribution of the sums

13. Example 7.7(a): minutes of app engagement

Fill the middle

The mean number of minutes of app engagement by a tablet user is 8.2 minutes, with a standard deviation of one minute. Take a sample of size 70. Fill in the mean and standard deviation of the sums, to two decimal places where needed.

Fill in the blanks

\mu_574 = (70)(8.2) = 8.37, \qquad \sigma____ = (\sqrt___)(1) \approx ___

Why: The centre multiplies by the sample size: 70 times 8.2 is 574 minutes. The spread multiplies by the square root of the sample size: the square root of 70 is about 8.37, and times one minute that is 8.37 minutes.

Both answers are in minutes, the units of a single session, because a total of minutes is still minutes. A total of 70 sessions is centred at 574 minutes, a little under ten hours, and typically lands within about 8 minutes of that.

If you wrote 70 for the standard deviation, you multiplied σ by n. That would say the total of 70 sessions is spread over hundreds of minutes either side of 574, which the dice already showed is far too wide.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.7(a)

14. Trap: multiplying σ by n

Trap

The trap

Example 7.5: find P(ΣX > 7500) when the population has mean 90 and standard deviation 15 and n is 80. A first attempt scales both parameters by 80:

\[ \sigma_{\Sigma X} = (80)(15) = 1200 \]

\[ z = \frac{7500 - 7200}{1200} = 0.25 \]

\[ P \approx 0.40 \]

Wrong. It says four samples in ten total more than 7,500.

The fix

The centre scales by n, but the spread scales only by the square root of n.

\[ \sigma_{\Sigma X} = (\sqrt{80})(15) \approx 134.16 \]

\[ z = \frac{300}{134.16} \approx 2.24 \]

\[ P \approx 0.0127 \]

The wrong answer is thirty times too big. Multiplying σ by n would be right only if all 80 values rose and fell together; Part 2 shows why independent values do not. As a quick check, a spread of 1,200 on a total of 7,200 would make totals near 6,000 or 8,400 routine, which the dice have already shown does not happen.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 371-372 — Example 7.5(a)

15. A mean, a sum, or one value?

Sorting

Section 7.3 gives the rule for choosing a distribution: if you are asked about a mean, use the theorem for means; about a sum or total, use the theorem for sums; about an individual value, use neither, but the distribution of X itself. Sort these questions.

Sort into buckets

Which distribution does each question need?

theorem for means
the probability that the average age of 50 iPad users is above 36; the 90th percentile of the mean stress score of 75 students
theorem for sums
the probability that the ages of 50 users add to more than 1,800 years; the 95th percentile of the total minutes of 70 sessions; whether 34 cans hold more than 544.34 ounces between them
distribution of one value
the probability that one randomly chosen user is older than 40; the chance a single can holds more than 16.2 ounces
mean
Words like average and mean point to the sample mean, whose spread is σ divided by the square root of n.
sum
Words like total, add to, between them and altogether point to the sum, whose spread is σ multiplied by the square root of n.
one
A question about a single randomly chosen value is about X itself. No theorem applies, and you need the population's own distribution.

Two items look almost the same: the average age of 50 users above 36, and the ages of 50 users adding to more than 1,800 years. They are in fact the same event, since 1,800 is 50 times 36, and Part 5 shows the two theorems give the same probability for it. What differs is the ruler: years of average, or years of total.

The words carry the decision. Total, combined, add to and between them mean a sum; average and mean mean the mean; one randomly chosen means a single value. Make this choice before any arithmetic, because each bucket uses a different standard deviation.

The single-value questions are the ones to be careful with. The central limit theorem says nothing about one value, so if the population is skewed or unknown, a normal calculation for one can or one user is simply not justified.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.3, pp. 373-374 — when to use the central limit theorem

16. Why the two rates differ

Section

Part 2

17. The centre adds up in full

Concept

Start with the easy half. Each value in the sample has the same mean, μX. The mean of a sum is the sum of the means, so adding n values adds n copies of μX.

\[ \mu_{\Sigma X} = \mu_{X_1} + \mu_{X_2} + \cdots + \mu_{X_n} \]

\[ \mu_{\Sigma X} = \mu_X + \mu_X + \cdots + \mu_X \]

\[ \mu_{\Sigma X} = n\mu_X \]

There is nothing to cancel here. Every value contributes its full average to the total, whatever the others do, so the centre grows in exact proportion to n: twice the sample, twice the expected total.

For the dice, each die averages 3.5, so ten dice average 35 and a hundred dice average 350. For the iPad users of Example 7.6, each age averages 34 years, so fifty ages average 1,700 years. The centre is always the natural guess, and it is right.

Keep this line beside the next slide. The spread does not behave like this, and the reason is that deviations, unlike averages, carry signs.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-371 — the mean of the sums is the original mean multiplied by the sample size

18. Two dice: high and low rolls cancel

Concept

Now the spread. Take the smallest possible total, two dice, and list all 36 equally likely outcomes. Each cell shows the sum.

Figure (svg): A six by six grid of the 36 equally likely outcomes of two dice, each cell showing the sum; the anti-diagonal of sevens is highlighted. To the right, a bar chart of how many cells give each sum: 1, 2, 3, 4, 5, 6, 5, 4, 3, 2, 1 for sums 2 to 12.

Extreme sums need both dice to be extreme together; middle sums can be made many ways, because a high roll and a low roll cancel. That cancelling is why the spread of a sum grows slower than the sum.

To get a sum of 12, both dice must show six: one cell out of 36. To get a sum of 7 there are six cells, because a high roll on one die can be balanced by a low roll on the other. Extremes need agreement; the middle can be reached many ways.

If the spreads simply added, the sum of two dice would have standard deviation 1.708 plus 1.708, about 3.42. Compute it from the 36 cells instead and you get something smaller.

\[ \sigma_{X_1 + X_2} = \sqrt{\tfrac{35}{6}} \approx 2.415 \quad < \quad 1.708 + 1.708 = 3.416 \]

\[ 2.415 = \sqrt{2}\,(1.708) \]

The spread of two dice is the square root of 2 times the spread of one. That is the theorem's rule at n equal to 2, and the grid shows its cause: independent departures from the mean partly cancel, so they do not pile up in full.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 371 — standard deviation of the sums: the original multiplied by root n

19. Where the square root comes from

Concept

What does add up in full, for independent values, is the variance, the square of the standard deviation. That is a result from probability you can take on trust here; the two dice just confirmed it, since 35 over 12 plus 35 over 12 is 35 over 6.

\[ \sigma^2_{\Sigma X} = \sigma^2_{X_1} + \sigma^2_{X_2} + \cdots + \sigma^2_{X_n} \]

Each value has the same variance, σX squared, so there are n equal terms.

\[ \sigma^2_{\Sigma X} = n\,\sigma_X^2 \]

The standard deviation is the square root of the variance. Take the square root of both sides, and the square root splits over the product.

\[ \sigma_{\Sigma X} = \sqrt{n\,\sigma_X^2} \]

\[ \sigma_{\Sigma X} = \sqrt{n}\,\sigma_X \]

So the square root is not a rule to memorise; it is what is left after squaring the spreads, adding them, and un-squaring. The same three moves, with a division by n squared for an average, give the σ over root n of Section 7.1.

And the whole argument rests on the word independent. If the values moved together, the variances would not simply add, and the square root would be lost; the boundary slide later in this part pushes on exactly that.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-371 — the sum's standard deviation

20. A straight line and a slow curve

Picture it

Put the two rates on one graph. Take a population with mean 15 and standard deviation 15, so single values are wildly variable, and plot the total's centre and the total's spread against n.

Figure (svg): Against sample size n from 0 to 100, a straight blue line 15n rising to 1,500 and an orange curve 15 times the square root of n rising slowly to 150. Dots mark 375 and 75 at n equal to 25, and 1,500 and 150 at n equal to 100.

The centre grows in a straight line; the spread grows too, but bends over. At every n the gap between them widens.

The blue line is the centre, 15 times n: straight, because it is proportional to n. The orange curve is the spread, 15 times the square root of n: it keeps rising, but it bends over, because to double the square root you must quadruple n.

\[ n = 25: \quad 375 \text{ vs } 75; \qquad n = 100: \quad 1500 \text{ vs } 150 \]

At n equal to 25 the centre is five times the spread; at n equal to 100 it is ten times. The gap between the two curves is the practical content of the section: the total becomes less and less uncertain relative to its size, even though its uncertainty in absolute terms never stops growing.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 371 — mean multiplied by n, standard deviation by root n

21. The same population summed at four scales

Worked example

For the population with mean 15 and standard deviation 15, compute the spread of the total as a share of its centre at n equal to 1, 4, 25 and 100. Call that share the relative spread.

\[ \frac{\sigma_{\Sigma X}}{\mu_{\Sigma X}} = \frac{\sqrt{n}\,(15)}{n(15)} \]

n = 1

Why: one value: spread 15 on a centre of 15.

\[ \frac{15}{15} = 1 = 100\% \]

n = 4

Why: spread doubles to 30; centre quadruples to 60.

\[ \frac{(2)(15)}{(4)(15)} = \frac{30}{60} = 50\% \]

n = 25

Why: spread 75, centre 375.

\[ \frac{(5)(15)}{(25)(15)} = \frac{75}{375} = 20\% \]

n = 100

Why: spread 150, centre 1,500.

\[ \frac{(10)(15)}{(100)(15)} = \frac{150}{1500} = 10\% \]

Figure (svg): Five orange bars showing the standard deviation of a total divided by its mean: 100 percent at n equal to 1, 50 percent at 4, 20 percent at 25, 10 percent at 100 and 5 percent at 400.

The absolute spread of the total rises with n, but the centre rises faster, so the spread as a share of the total keeps falling: halving every time n is multiplied by four.

Look at the pattern in the four lines before reading the chart. The top of each fraction grows like the square root of n, the bottom like n itself, so the fraction shrinks like one over the square root of n. The fifth bar, at n equal to 400, continues the pattern: 20 times 15 over 400 times 15 is one twentieth, 5 percent.

Every time n is multiplied by four, the relative spread halves. The absolute spread went from 15 to 150, a tenfold rise, and at the same time the total became ten times more predictable in proportion. Both statements are true; they measure different things.

Check with the simplified formula

Why: cancel one root n from top and bottom: the share is σ over root n times μ, so at n = 100 it is one tenth.

\[ \frac{\sqrt{n}\,\sigma}{n\mu} = \frac{\sigma}{\sqrt{n}\,\mu} = \frac{15}{(10)(15)} = 0.10 \]

22. Decoding the relative spread

Notation

The ratio on the last slide is worth one formula of its own, because it is how actuaries, engineers and anyone who adds up many uncertain things think about totals.

Annotate

On: \( \frac{\sigma_{\Sigma X}}{\mu_{\Sigma X}} = \frac{\sigma_X}{\sqrt{n}\,\mu_X} \)

  • The spread of the total as a fraction of its centre: how uncertain the total is, relative to its size.
  • The same fraction for one value. For the dice it is 1.708 over 3.5, about 0.49.
  • The only place n appears. The relative spread of a total falls like one over the square root of n.

For ten dice the relative spread is 0.49 divided by the square root of 10, about 0.154: the total of ten dice is typically within 15 percent of 35. For a hundred dice it is 0.049, within about 5 percent of 350.

\[ \frac{5.40}{35} \approx 0.154, \qquad \frac{17.08}{350} \approx 0.049 \]

Compare this with Section 7.1: σX over the square root of n is the spread of the average. Divide it by μX and you get the same ratio. That is no accident, since the total is just n times the average, and the relative spread ignores the scaling by n.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-371 — the sum's mean and standard deviation

23. Slide n and watch the total

Tweak it

The slider draws the distribution of the total of n waiting times, each with mean 1 minute and standard deviation 1 minute, so the total follows N(n, the square root of n). Move n and watch the centre and the width separately.

Parameter explorer

Set n to 4, then 16, then 64. Each time, read the centre and estimate the width of the curve. By how much does the centre move each time, and by how much does the width grow?

\[ \Sigma X \sim N\big({n},\ \sqrt{{n}}\big) \]

  • n — from 1 to 100: number of waiting times n

\[ n = 4 \Rightarrow N(4, 2), \quad n = 16 \Rightarrow N(16, 4), \quad n = 64 \Rightarrow N(64, 8) \]

Try one more setting. At n equal to 100 the curve is centred at 100 with standard deviation 10: most totals fall between 80 and 120, within 20 percent of the centre. At n equal to 1 a single wait could easily be 0 or 3, a relative spread of 100 percent.

Each quadrupling of n moves the centre four times further out but only doubles the width. Notice too that the peak drops as the curve widens, because the area under it must stay one.

One honest caveat: waiting times are skewed, so at n equal to 1 or 2 the true distribution of the total is not the symmetric curve drawn here. The slider shows the theorem's normal curve; it becomes a good description only as n grows.

24. Ten thousand values from a wild population

Estimation

Single values from the population with mean 15 and standard deviation 15 are wildly variable: the typical value is off by as much as its own size. Now add up ten thousand of them. Estimate before you compute.

Predict first

For the total of 10,000 values, roughly how large is the standard deviation as a fraction of the total's mean?

  • about 1 percent
  • about 10 percent
  • about 100 percent, the same as one value
  • about 0.01 percent

Correct: about 1 percent

Why: The relative spread is the single value's relative spread, which is one, divided by the square root of n. The square root of ten thousand is one hundred, so the total's spread is about one hundredth of its mean.

\[ \frac{\sigma_X}{\sqrt{n}\,\mu_X} = \frac{15}{(100)(15)} = 0.01 \]

In numbers: the total is centred at 150,000 with standard deviation 1,500. A population in which single values are as uncertain as can be gives a total you can predict to within about 1 percent.

The option at 0.01 percent divides by n instead of its square root. The option at 10 percent is what you get at n equal to 100, so it answers a smaller question. Ten thousand is a hundred squared, so the square root of n is one hundred.

25. Four spreads for Example 7.6's sum of 50 ages

Error analysis

Example 7.6: ages of iPad users have mean 34 years and standard deviation 15 years, and the sample size is 50. Four students computed the standard deviation of the sum of the 50 ages. One is right. Name the fault in each of the others.

Annotate

On: \( (50)(15) \quad \frac{15}{\sqrt{50}} \quad \sqrt{(50)(15)} \quad (\sqrt{50})(15) \)

  • Multiplies σ by n. That treats the fifty ages as moving together, so their spreads add in full; independent values partly cancel.
  • The standard error of the mean, from Section 7.1. A correct formula for the wrong question: it is the spread of the average age, not the total.
  • Puts σ under the square root as well. Only n belongs under the root; σ multiplies it from outside.
  • Correct: the square root of the sample size times the population standard deviation.

The standard error answer is the dangerous one, because nothing about it looks wrong: it is the right formula for a mean. The scale check catches it. A total of 50 ages sits near 1,700 years, and a spread of 2 years on a total of 1,700 is absurdly tight.

Use the same scale check on the others. A spread of 750 on a centre of 1,700 is enormous, 44 percent, as uncertain as a single age; 27 is too small; 106 is about 6 percent, which is what the relative-spread formula predicts, 15 over 34 divided by the square root of 50.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 372-373 — Example 7.6(a)

26. What if the values are not independent?

Edge cases

The square root came from adding variances, which needs independence. Push on that. Suppose the 80 values of Example 7.5 are not independent at all: they all move together, so if one is 15 above the mean, every one is.

Figure (svg): Two curves centred at 7,200. A tall narrow blue curve for independent values, standard deviation 134.16, and a low wide red curve for values that all move together, standard deviation 1,200, spreading from about 3,600 to 10,800.

Same centre, very different risk. The square-root law is a gift of independence: take it away and the spread of the total grows in full proportion to n.

Discussion prompt

If all 80 values rise and fall together, what is the standard deviation of their total, and what is the probability that the total exceeds 7,500 now? What does this tell you about pooling risks that are linked, such as flood damage to houses on the same street?

This is why the trap on multiplying σ by n is a trap and not a harmless alternative: it is the correct answer to a different situation, perfect dependence. The theorem's answer, the square root of n, is right only when the values are independent.

Real data usually sit between the two extremes, and then neither formula is exact. That is one reason the book insists on random samples: random sampling is how you earn the independence the square root needs.

27. Why insurance works

Real world

An insurer sells car policies. For one policy, the claim in a year has mean 400 dollars and standard deviation 2,000 dollars: usually nothing, occasionally a great deal. One policy on its own is close to a coin toss with a large downside.

Discussion prompt

The insurer holds 10,000 independent policies. Find the mean and standard deviation of the total claims, and the spread as a share of the mean. Then explain in a sentence why the insurer can safely charge each customer about 450 dollars.

One policy has a relative spread of 5, as uncertain as a thing can be; ten thousand of them have 0.05. That hundredfold drop is the square root of ten thousand, and it is the whole business model of insurance, pension funds and casinos.

The boundary slide says where the model breaks. Independent car accidents pool nicely; a hurricane, a pandemic or a market crash hits many policies at once, and then the square root law no longer protects the pool.

28. Probabilities for a sum

Section

Part 3

29. Every probability for a sum is an area

Concept

Once you have written ΣX ~ N(nμX, the square root of n times σX), a probability for the sum is an ordinary normal-curve question. It is the area under that curve over the values the question describes.

\[ P(\Sigma X > \Sigma x) = \text{area to the right of } \Sigma x \text{ under } N\big(n\mu_X,\ \sqrt{n}\,\sigma_X\big) \]

So the routine is the one from Chapter 6, with one new line at the start. Write the distribution of the sum. Draw the curve, mark the value or values in the question, and shade the area asked for. Then find that area, with a z-score and a table, or directly with a calculator.

  1. Distribution: centre n times μX, spread root n times σX.
  2. Picture: mark the value, shade the side the question asks about.
  3. Standardise: how many standard deviations of the sum is the value from the centre?
  4. Area: from the table or calculator, then a sentence in context.

Three shapes of question cover almost everything: more than a value, a right tail; less than a value, a left tail; between two values, a band. For a continuous total, more than and at least give the same area, so you need not worry about whether the endpoint is included.

Draw the picture every time, even when the calculator will do the area. The picture tells you whether the answer should be tiny, near a half, or near one, and that single glance catches most wrong answers before they are written down.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 371 — to find probabilities for sums

30. The z-score for a sum

Concept

Any z-score is a value minus its centre, divided by its spread. For a sum, the value is one observed total, Σx, and the centre and spread are those of the sum.

\[ z = \frac{\text{value} - \text{centre}}{\text{spread}} \]

\[ z = \frac{\Sigma x - \mu_{\Sigma X}}{\sigma_{\Sigma X}} \]

Now substitute the theorem's centre and spread. The numerator measures how far the total is from what n values should add to; the denominator is the size of one standard deviation of the total.

\[ z = \frac{\Sigma x - (n)(\mu_X)}{(\sqrt{n})(\sigma_X)} \]

Compare it with the z-score for a mean from Section 7.1. The shape is identical, value minus centre over spread, but every piece is the version for the total: a total instead of a mean on top, n times μX instead of μX, and the square root of n times σX instead of σX over the square root of n.

This is the formula the book gives for the sum. Its z is read exactly like any other: 2 means two standard deviations of the total above the centre, minus 1 means one standard deviation below.

The calculator route in the book, normalcdf with the lower value, the upper value, n times the mean, and the square root of n times the standard deviation, is doing this standardising for you. If you work from a table, you do it by hand, and the arithmetic on these slides shows both.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 371 — the z-score associated with ΣX

31. Decoding the z-score for a sum

Notation

Four quantities, and each belongs to a different level: one particular total, the sample size, and two population parameters.

Annotate

On: \( z = \frac{\Sigma x - (n)(\mu_X)}{(\sqrt{n})(\sigma_X)} \)

  • One particular sum, the value in the question, such as 7,500. Lower-case: an observed number, not the random variable.
  • The mean of the sums: where totals of n values centre.
  • How far this total is from the centre, in the original units, with a sign.
  • One standard deviation of the sums. Dividing by it counts standard deviations of the total, not of one value.

The most common slip is to use the right numerator with the wrong denominator: σX alone, or σX over the square root of n. Both give a z far too large, and a probability that is far too extreme.

\[ \text{Example 7.5: } \quad z = \frac{7500 - (80)(90)}{(\sqrt{80})(15)} = \frac{300}{134.16} \approx 2.24 \]

Say what the number means before moving on: 7,500 is about two and a quarter standard deviations of the total above the expected total of 7,200.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 371 — Σx is one sum; (n)(μX) is the mean of ΣX

32. Example 7.5(a): a total of more than 7,500

Worked example

An unknown distribution has a mean of 90 and a standard deviation of 15. A sample of size 80 is drawn randomly. Find the probability that the sum of the 80 values is more than 7,500.

\[ \Sigma X \sim N(7200,\ 134.16) \]

Write the event

Why: more than 7,500, so the area to the right.

\[ P(\Sigma X > 7500) \]

Standardise: substitute

Why: the sum's centre and spread.

\[ z = \frac{7500 - 7200}{134.16} \]

Subtract

Why: 7,500 is 300 above the centre.

\[ z = \frac{300}{134.16} \]

Divide

Why: about two and a quarter standard deviations.

\[ z \approx 2.236 \]

Area to the right

Why: one minus the area to the left of z.

\[ P(Z > 2.236) = 1 - 0.9873 = 0.0127 \]

Figure (svg): The normal curve for the sum of 80 values, centred at 7,200, with the sliver of area to the right of 7,500 shaded yellow and labelled 0.0127. A second row under the axis gives z-scores; 7,500 sits at z equal to 2.24.

The shaded tail is the probability. 7,500 is 300 above the centre, which is 2.24 standard deviations of the sum, so the tail is thin.

In context: the probability that the total of 80 randomly drawn values exceeds 7,500 is about 0.0127, so only about one sample in eighty has a total that large. The picture shows why: 7,500 sits well out in the right tail.

Look at the size of the step you took. For one value, 300 would be twenty standard deviations of 15; for the total it is only 2.24, because the relevant ruler is the spread of the total, 134.16. Getting that ruler right is the whole difference between a sensible answer and nonsense.

With a calculator the whole thing is one line, normalcdf of 7500, 1E99, 80 times 90, the square root of 80 times 15, which the book reports as 0.0127. The 1E99 is just a very large upper limit, standing in for infinity.

Check with the exact z

Why: 7,500 minus 7,200 is 300, and 300 over root 80 times 15 is exactly root 5; the area beyond root 5 matches the book's 0.0127.

\[ z = \frac{300}{15\sqrt{80}} = \sqrt{5} \approx 2.2361, \quad 1 - \Phi(2.2361) = 0.0127 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 371-372 — Example 7.5(a), Figure 7.3

33. Try It 7.5: a total of more than 2,400

Worked example

An unknown distribution has a mean of 45 and a standard deviation of eight. A sample of size 50 is drawn randomly. Find the probability that the sum of the 50 values is more than 2,400.

\[ \mu_X = 45, \quad \sigma_X = 8, \quad n = 50 \]

Centre of the sum

Why: 50 values averaging 45.

\[ \mu_{\Sigma X} = (50)(45) = 2250 \]

Spread of the sum

Why: root 50 is about 7.071.

\[ \sigma_{\Sigma X} = (\sqrt{50})(8) \approx 56.57 \]

Standardise

Why: 2,400 is 150 above the centre.

\[ z = \frac{2400 - 2250}{56.57} = \frac{150}{56.57} \]

Divide

Why: more than two and a half standard deviations.

\[ z \approx 2.652 \]

Area to the right

Why: one minus the area to the left.

\[ P(\Sigma X > 2400) = 1 - 0.996 = 0.004 \]

Figure (svg): The normal curve for the sum of 50 values, centred at 2,250 with standard deviation 56.57, and the tiny area to the right of 2,400 shaded yellow. The z row puts 2,400 at 2.65.

2,400 is 150 above the centre: two and two thirds standard deviations of the sum. Only about 4 samples in 1,000 total that much.

In context: there is only about a 0.4 percent chance that 50 values drawn from this population add to more than 2,400. The book's answer is 0.0040.

Notice how little room 150 leaves. For one value, 150 above the mean would be almost 19 standard deviations of 8, unthinkable; for the total it is 2.65 standard deviations of 56.57, rare but possible. The spread of the total is what sets the scale.

Check through the mean

Why: a total over 2,400 is a mean over 48; the mean has spread 8 over root 50, and the z-score must match.

\[ \frac{48 - 45}{8/\sqrt{50}} = \frac{3}{1.1314} \approx 2.652 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 372 — Try It 7.5

34. Example 7.6(a)(b): fifty ages between 1,500 and 1,800 years

Worked example

The mean age of iPad users is 34 years, with standard deviation 15 years. The sample size is 50. Give the distribution of the sum of the ages, then find the probability that the sum is between 1,500 and 1,800 years.

(a) Centre of the sum

Why: fifty ages averaging 34 years.

\[ \mu_{\Sigma X} = (50)(34) = 1700 \]

Spread of the sum

Why: root 50 times 15.

\[ \sigma_{\Sigma X} = (\sqrt{50})(15) \approx 106.07 \]

Distribution

Why: normal for sums, by the theorem.

\[ \Sigma X \sim N(1700,\ 106.07) \]

Part (b) asks about a band, so there are two values to standardise. Before computing, place them on the picture in your head: 1,500 is 200 below 1,700, almost two standard deviations of the sum; 1,800 is only 100 above it. Expect most, but not all, of the area.

(b) Standardise the lower end

Why: 1,500 is 200 below the centre.

\[ z_1 = \frac{1500 - 1700}{106.07} \approx -1.886 \]

Standardise the upper end

Why: 1,800 is 100 above the centre.

\[ z_2 = \frac{1800 - 1700}{106.07} \approx 0.943 \]

Area between

Why: area left of the upper z minus area left of the lower z.

\[ P = 0.8271 - 0.0297 = 0.7974 \]

Figure (svg): The normal curve for the sum of 50 ages, centred at 1,700, with the region from 1,500 to 1,800 shaded blue and labelled 0.7974. The z row shows 1,500 at minus 1.89 and 1,800 at 0.94.

The band is lopsided about the centre: it reaches almost two standard deviations to the left but less than one to the right, so the area is neither 68 nor 95 percent.

In context: there is about an 80 percent chance that the ages of 50 randomly chosen iPad users add up to between 1,500 and 1,800 years. The book's calculator line, normalcdf of 1500, 1800, 50 times 34, root 50 times 15, gives 0.7974.

The band is not symmetric about 1,700, so do not reach for the Empirical Rule. It stretches almost two standard deviations to the left but under one to the right, which is why the answer, 0.80, lies between 68 and 95 percent.

Check the pieces add up

Why: the three regions, left tail, band and right tail, must total one.

\[ 0.0297 + 0.7974 + 0.1729 = 1 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 372-373 — Example 7.6(a)(b)

35. Try It 7.6: thirty-nine ages

Worked example

Now the mean age is 35 years with standard deviation 10 years, and the sample size is 39. Find the distribution of the sum, the probability that it is between 1,400 and 1,500 years, and its 90th percentile.

(a) Centre and spread

Why: 39 times 35, and root 39 times 10.

\[ \mu_{\Sigma X} = (39)(35) = 1365, \quad \sigma_{\Sigma X} = (\sqrt{39})(10) \approx 62.45 \]

Distribution

Why: normal for sums.

\[ \Sigma X \sim N(1365,\ 62.45) \]

The distribution is centred at 1,365, so both ends of the band in part (b), 1,400 and 1,500, lie above the centre. Expect a band on the right shoulder of the curve and a probability under one half. The procedure is the same as Example 7.6; only the numbers change.

(b) Standardise both ends

Why: both lie above the centre.

\[ z_1 = \frac{1400 - 1365}{62.45} \approx 0.560, \quad z_2 = \frac{1500 - 1365}{62.45} \approx 2.162 \]

Area between

Why: difference of the two left areas.

\[ P = 0.9847 - 0.7124 = 0.2723 \]

(c) The 90th percentile

Why: the z with 0.90 to its left is 1.2816.

\[ k = 1365 + (1.2816)(62.45) \approx 1445 \]

Figure (svg): The normal curve for the sum of 39 ages, centred at 1,365, with the region from 1,400 to 1,500 shaded and labelled 0.2723, and a pink line at the 90th percentile, 1,445.0.

The whole shaded band lies to the right of the centre, so it must hold less than half the area. The 90th percentile falls inside it.

In context: there is about a 27 percent chance that the ages of 39 users add to between 1,400 and 1,500 years, and 90 percent of such totals are at most about 1,445 years. Part 4 explains the percentile step in full; here it just shows where it lands.

The whole band lies to the right of the centre, so the probability had to be under a half. And the 90th percentile, 1.28 standard deviations up, falls inside the band, between 1,400 and 1,500.

Check the percentile by standardising it

Why: the answer should sit 1.2816 standard deviations of the sum above 1,365.

\[ \frac{1445.03 - 1365}{62.45} \approx 1.2816 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Try It 7.6

36. One value over 93.75, or eighty values over 7,500?

Prediction

Stay with Example 7.5's population, mean 90 and standard deviation 15, and suppose for this slide that it is normal, so one value has a known distribution too. The total of 80 values exceeds 7,500 exactly when their average exceeds 93.75. Compare two events.

Predict first

Which is more likely: one randomly drawn value exceeds 93.75, or the total of 80 values exceeds 7,500?

  • one value exceeding 93.75 is far more likely
  • the two are equally likely, since 7,500 is 80 times 93.75
  • the total exceeding 7,500 is far more likely
  • it cannot be decided without the sample data

Correct: one value exceeding 93.75 is far more likely

Why: For one value, 93.75 is only a quarter of a standard deviation above 90, so about 40 percent of values exceed it. For the total, 7,500 is 2.24 standard deviations of the sum above 7,200, so only about 1.3 percent of totals do.

\[ P(X > 93.75) = P(Z > 0.25) \approx 0.40 \]

\[ P(\Sigma X > 7500) = P(Z > 2.24) \approx 0.0127 \]

Scaling the threshold by 80 does not scale the probability. A single value above average is common; eighty values that are above average on the whole is rare, because the high ones and the low ones cancel. That is the spread of a total growing only like the square root of n, felt as a probability.

It also shows why Section 7.3 insists on asking whether a question is about one value or about a total: the same-looking numbers give answers thirty times apart.

37. Example 7.7(c): at least ten hours of engagement

Worked example

App engagement per session has mean 8.2 minutes and standard deviation one minute, and n is 70. Find the probability that the sum of the sample is at least ten hours.

\[ \Sigma X \sim N(574,\ 8.37) \text{ in minutes} \]

Convert the units

Why: the distribution is in minutes, so ten hours must be too.

\[ 10 \text{ hours} = (10)(60) = 600 \text{ minutes} \]

Write the event

Why: at least 600: the area to the right.

\[ P(\Sigma X \ge 600) \]

Standardise

Why: 600 is 26 minutes above the centre.

\[ z = \frac{600 - 574}{8.37} = \frac{26}{8.37} \]

Divide

Why: over three standard deviations of the total.

\[ z \approx 3.108 \]

Area to the right

Why: a far-tail area.

\[ P(Z \ge 3.108) \approx 0.0009 \]

Figure (svg): The normal curve for the total of 70 app sessions, centred at 574 minutes with standard deviation 8.37. A yellow line at 600 minutes sits where the curve has almost reached the axis; the area to its right is 0.0009. The z row puts 600 at 3.11.

Ten hours is 26 minutes above the centre, more than three standard deviations of the total: the shaded area is too thin to see.

In context: there is only about a 0.09 percent chance, roughly one in a thousand, that 70 randomly chosen sessions add up to ten hours or more. For a continuous total, at least and more than give the same area, so the equals sign does not matter here.

The unit conversion is the step that decides everything. The picture's axis is in minutes; if 10 were marked on it, the mark would sit hundreds of minutes to the left of the curve, and that is the trap on the next slide.

Check with the book's calculator line

Why: normalcdf of 600, E99, 70 times 8.2, root 70 times 1 is printed as 0.0009, which is our 0.00094 rounded.

\[ 1 - \Phi(3.1076) = 0.00094 \approx 0.0009 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.7(c)

38. Trap: ten hours typed in as 10

Trap

The trap

Example 7.7(c) asks for the probability that 70 sessions total at least ten hours. A first attempt puts the 10 straight in:

\[ z = \frac{10 - 574}{8.37} \]

\[ z \approx -67.4 \]

\[ P(\Sigma X \ge 10) \approx 1 \]

Wrong. The question is in hours; the distribution is in minutes.

The fix

Convert the question's value into the units of the distribution first.

\[ 10 \text{ h} = 600 \text{ min} \]

\[ z = \frac{600 - 574}{8.37} \approx 3.11 \]

\[ P \approx 0.0009 \]

The wrong answer says a total of ten hours is certain; the right one says it is a one-in-a-thousand event. A z-score of minus 67 is itself the warning: no real question lives 67 standard deviations from the centre. When a z is that extreme, look for a units mismatch before you believe it.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.7(c): ten hours = 600 minutes

39. Try It 7.7(a): between seven and ten hours

Worked example

Same app sessions, n equal to 70. Find the probability that the sum of the sample is between seven hours and ten hours, and say what it means in context.

\[ \Sigma X \sim N(574,\ 8.37) \text{ in minutes} \]

Convert both ends

Why: seven and ten hours into minutes.

\[ 7 \text{ h} = 420 \text{ min}, \qquad 10 \text{ h} = 600 \text{ min} \]

Standardise the lower end

Why: 420 is 154 minutes below 574.

\[ z_1 = \frac{420 - 574}{8.37} \approx -18.41 \]

Standardise the upper end

Why: as in Example 7.7(c).

\[ z_2 = \frac{600 - 574}{8.37} \approx 3.108 \]

Area between

Why: the area left of minus 18.41 is zero to any precision.

\[ P = 0.9991 - 0 = 0.9991 \]

Figure (svg): On an axis from 400 to 620 minutes, the curve for the total of 70 sessions is a narrow spike at 574. Yellow lines at 420 and 600 enclose all but a sliver of it; the enclosed area is 0.9991.

Seven hours is so far below the centre, over 18 standard deviations, that the left end of the band cuts off nothing. Only the right-hand tail beyond 600 is missing.

In context: it is almost certain, a probability of about 0.999, that the total engagement time of 70 sessions lies between seven and ten hours. The only way to miss is on the high side, beyond ten hours, and that has the one-in-a-thousand chance you found before.

Here a huge z is genuine, not a units slip: seven hours really is 18 standard deviations of the total below the centre. The difference from the trap is that you converted first and the answer still makes sense in context.

Check with the complement

Why: the band misses only the right tail beyond 600, so the answer is one minus Example 7.7(c)'s tail.

\[ 1 - 0.0009 = 0.9991 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Try It 7.7(a)

40. Example 7.8: what do you need before any z-score?

Step zero

A campus stress study: stress scores follow a uniform distribution from one to five. For 75 students, the questions ask for the probability that the total of the 75 scores is less than 200, and for its 90th percentile. Before any arithmetic, decide what you need and where it comes from.

Discussion prompt

The question gives you a uniform distribution, not a mean and standard deviation. What two numbers does the theorem for sums need, how do you get them from U(1, 5), and what are the centre and spread of the total of 75 scores?

The first move is to recognise which kind of question this is: total of the 75 stress scores means the theorem for sums. The second is to notice that you are handed a shape, not parameters, so you must compute μX and σX yourself before the theorem can start.

The spread comes out exactly 10 because 75 times 16 over 12 is exactly 100. Clean numbers like that are a sign you have the formula right.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.3, p. 374 — Example 7.8: X ~ U(1, 5), μX = 3

41. Example 7.8(c)(d): the total stress score of 75 students

Worked example

Stress scores are uniform from one to five, and a sample of 75 students is taken. Find the probability that the total of the 75 scores is less than 200, and the 90th percentile of the total.

\[ \Sigma X \sim N(225,\ 10) \]

The step-zero slide already did the hard part: turning a uniform shape into a mean of 3 and a standard deviation of about 1.155, and those into a total centred at 225 with spread exactly 10. From here on, it is an ordinary normal question.

(c) Write the event

Why: less than 200: the area to the left.

\[ P(\Sigma X < 200) \]

Standardise

Why: 200 is 25 below the centre.

\[ z = \frac{200 - 225}{10} = -2.5 \]

Area to the left

Why: from the table or calculator.

\[ P(Z < -2.5) = 0.0062 \]

(d) The 90th percentile

Why: the z with 0.90 to its left is 1.2816.

\[ k = 225 + (1.2816)(10) \]

Multiply and add

Why: 12.8 points above the centre.

\[ k = 225 + 12.816 = 237.8 \]

Figure (svg): Left: a flat purple rectangle from 1 to 5 of height one quarter, the uniform distribution of one stress score, with mean 3 and standard deviation 1.155. An arrow labelled add 75 leads right to a bell curve centred at 225 with standard deviation 10. Its left tail below 200 is shaded red, area 0.0062, and a pink line marks the 90th percentile at 237.8.

A flat population, a bell-shaped total. The theorem does not care what the population looks like; it needs only the population's mean and standard deviation and a large enough sample.

In context: there is only about a 0.6 percent chance that 75 students' stress scores total less than 200, and 90 percent of such totals are at most about 237.8. A total of 200 would mean an average score of 2.67, and with 75 students that is well out of the ordinary.

Look at what the picture says about the theorem. Not one student's score is bell-shaped, yet the total of 75 is. The uniform population is the reason this example sits in the book: the theorem needed only μX, σX and a large n.

Check through the mean

Why: a total under 200 is a mean under 200 over 75; the mean has spread 1.155 over root 75, and the z must again be minus 2.5.

\[ \frac{2.667 - 3}{1.155/\sqrt{75}} = \frac{-0.3333}{0.1333} = -2.5 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.3, p. 374 — Example 7.8(c)(d); the answers are computed here

42. Percentiles for a sum

Section

Part 4

43. A percentile runs the z-score backwards

Concept

A probability question gives you a total and asks for an area. A percentile question gives you an area and asks for the total. So you run the z-score formula in reverse: find the z that has the right area to its left, then turn it back into a total.

\[ z = \frac{k - (n)(\mu_X)}{(\sqrt{n})(\sigma_X)} \]

Multiply both sides by the spread of the sum. The right side is now the distance of k from the centre, in the units of the problem.

\[ z(\sqrt{n})(\sigma_X) = k - (n)(\mu_X) \]

Add the centre to both sides to isolate k.

\[ k = (n)(\mu_X) + (z)(\sqrt{n})(\sigma_X) \]

Read that as a walk: start at the centre of the sums, then step z standard deviations of the sum. The z comes from the table or from invNorm; the book's calculator line, invNorm of the area to the left, n times the mean, root n times the standard deviation, does the whole walk at once.

The only change from Chapter 6 is which standard deviation you step with. It is the sum's, the square root of n times σX, never σX itself.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 372 — to find percentiles for sums

44. Example 7.5(b): the sum 1.5 standard deviations above the mean

Worked example

Same population as before: mean 90, standard deviation 15, n equal to 80. Find the sum that is 1.5 standard deviations above the mean of the sums.

\[ \Sigma X \sim N(7200,\ 134.16) \]

Set up the walk

Why: start at the centre, step 1.5 standard deviations of the sum.

\[ \Sigma x = (n)(\mu_X) + (z)(\sqrt{n})(\sigma_X) \]

Substitute

Why: z is given directly as 1.5.

\[ \Sigma x = (80)(90) + (1.5)(\sqrt{80})(15) \]

Multiply

Why: 1.5 standard deviations is 201.25.

\[ \Sigma x = 7200 + (1.5)(134.164) = 7200 + 201.246 \]

Add

Why: and round to one decimal place.

\[ \Sigma x = 7401.246 \approx 7401.2 \]

Figure (svg): The curve for the sum of 80 values centred at 7,200. Orange arrows step one full standard deviation of 134.16 and then half of one more, landing on a yellow line at 7,401.2; the z row there reads 1.5.

One and a half standard deviations of the sum, 201.25, added to the centre. The z row under the axis confirms where the line sits.

In context: a total of about 7,401.2 is one and a half standard deviations of the sum above the expected total of 7,200. The book's answer, 7,401.2, is 7,401.246 rounded.

Here the question hands you z, so no table is needed. That is worth noticing: a question that says so many standard deviations above the mean has done the percentile look-up for you, and only the walk is left.

Check by standardising the answer

Why: the total should sit 1.5 standard deviations of the sum above 7,200.

\[ \frac{7401.246 - 7200}{134.164} = \frac{201.246}{134.164} = 1.5 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 372 — Example 7.5(b)

45. Example 7.6(c): the 80th percentile of fifty ages

Worked example

Ages of iPad users: mean 34 years, standard deviation 15 years, n equal to 50, so the sum follows N(1700, 106.07). Find the 80th percentile for the sum of the 50 ages.

\[ \Sigma X \sim N(1700,\ 106.07) \]

A percentile question starts from an area, not a value. So the first move is the reverse look-up: search the body of the z-table for 0.80 and read off the z beside it, or let invNorm do it. Only then does the sum's distribution enter.

Find z

Why: the z with area 0.80 to its left, from the table or invNorm.

\[ P(Z < z) = 0.80 \Rightarrow z \approx 0.8416 \]

Set up the walk

Why: centre plus z standard deviations of the sum.

\[ k = 1700 + (0.8416)(106.07) \]

Multiply

Why: 0.84 standard deviations is about 89.3 years.

\[ k = 1700 + 89.27 \]

Add

Why: and round to one decimal place.

\[ k \approx 1789.3 \text{ years} \]

Figure (svg): The curve for the sum of 50 ages, centred at 1,700. Everything to the left of a pink line at 1,789.3 is shaded and labelled area 0.80; the unshaded right tail holds 0.20.

A percentile is read the other way round from a probability: fix the area on the left, 0.80, and find where the line must go. It sits 0.84 standard deviations above the centre.

In context: 80 percent of samples of 50 iPad users have ages that add up to at most about 1,789.3 years; the other 20 percent total more. That matches the book's invNorm line, invNorm of 0.80, 50 times 34, root 50 times 15.

Check the picture before the arithmetic. The 80th percentile must be above the centre, since more than half the area is to its left, but not far above, since 0.80 is not an extreme area. A bit under one standard deviation up is exactly that.

Check by finding the area to the left of the answer

Why: standardise 1,789.3 and the area to its left should come back as 0.80.

\[ \Phi\left(\frac{1789.27 - 1700}{106.07}\right) = \Phi(0.8416) = 0.8 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.6(c)

46. Example 7.7(b): the 95th percentile of 70 sessions

Worked example

App engagement: mean 8.2 minutes, standard deviation one minute, n equal to 70, so the sum follows N(574, 8.37). Find the 95th percentile for the sum of the sample and interpret it in a complete sentence.

\[ \Sigma X \sim N(574,\ 8.37) \]

The 95th percentile has 95 percent of the area to its left and only 5 percent to its right, so it sits well up the right shoulder of the curve. The z that cuts off that area is one worth knowing by heart, because it appears again in every later chapter: 1.645.

Find z

Why: 0.95 to the left.

\[ P(Z < z) = 0.95 \Rightarrow z \approx 1.645 \]

Set up the walk

Why: centre plus z standard deviations of the sum.

\[ k = 574 + (1.645)(8.3666) \]

Multiply

Why: about 13.76 minutes above the centre.

\[ k = 574 + 13.762 \]

Add

Why: to two decimal places.

\[ k \approx 587.76 \text{ minutes} \]

Figure (svg): The curve for the total of 70 sessions, centred at 574 minutes, shaded from the left up to a pink line at 587.76 minutes, labelled 0.95; the small right tail beyond it holds 0.05. The z row puts the line at 1.64.

Ninety-five percent of totals of 70 sessions fall at or below the pink line. It is 1.645 standard deviations of the total above 574.

The sentence the book asks for: ninety-five percent of the sums of app engagement times for samples of 70 sessions are at most 587.76 minutes. Equivalently, only 5 percent of such totals exceed 587.76 minutes, just under nine hours and 48 minutes.

Notice what the sentence is about. It is about totals of 70 sessions, not about single sessions: no statement here says anything about how long one person uses the app. Naming the random variable in the sentence is what keeps the interpretation honest.

Check by standardising the answer

Why: 587.76 should sit 1.645 standard deviations of the sum above 574.

\[ \frac{587.76 - 574}{8.3666} = \frac{13.76}{8.3666} \approx 1.645 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.7(b)

47. Try It 7.7(b): the 84th and 16th percentiles

Worked example

Same 70 sessions. Find the 84th and the 16th percentiles for the sum of the sample and interpret them in context.

\[ \Sigma X \sim N(574,\ 8.37) \]

Two percentiles, one look-up. Because the normal curve is symmetric, the z with 0.16 to its left is exactly the negative of the z with 0.84 to its left: 0.16 is the area in each tail beyond the pair. So find one z and use it with both signs.

Find the two z-values

Why: 0.84 and 0.16 to the left; symmetric about zero.

\[ z_{0.84} \approx 0.9945, \qquad z_{0.16} \approx -0.9945 \]

84th percentile: walk up

Why: about one standard deviation of the sum.

\[ k_{84} = 574 + (0.9945)(8.3666) = 574 + 8.32 \]

Add

Why: to two decimal places.

\[ k_{84} \approx 582.32 \]

16th percentile: walk down

Why: the same distance below.

\[ k_{16} = 574 - 8.32 \]

Subtract

Why: to two decimal places.

\[ k_{16} \approx 565.68 \]

Figure (svg): The curve for the total of 70 sessions. Pink lines at 565.68 and 582.32 minutes, labelled the 16th and 84th percentiles, enclose a shaded middle holding 0.68; each tail outside holds 0.16.

The 16th and 84th percentiles sit almost exactly one standard deviation of the total, 8.37 minutes, either side of 574: the Empirical Rule's middle 68 percent.

In context: 84 percent of totals of 70 sessions are at most 582.32 minutes, and 16 percent are at most 565.68 minutes. So the middle 68 percent of such totals lie between about 565.7 and 582.3 minutes, within about 8.3 minutes of 574.

These two percentiles were chosen to make a point. They sit almost exactly one standard deviation of the sum either side of the centre, which is the Empirical Rule's 68 percent band: you could have written them down from the rule, and the calculation confirms it.

Check the band against the Empirical Rule

Why: 574 plus and minus one standard deviation of the sum should nearly reproduce the two percentiles.

\[ 574 \pm 8.37 = 565.63 \text{ and } 582.37 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Try It 7.7(b)

48. Which z does each percentile need?

Matching

Every percentile calculation on these slides began by turning an area into a z-value. Match each percentile to its z, then to the total it gives for Example 7.7's sum, N(574, 8.37).

Match the pairs

  • a. 95th percentile
  • b. 90th percentile
  • c. 84th percentile
  • d. 50th percentile
  • e. 16th percentile
  • v. z = 1.645, total 587.76
  • w. z = 1.2816, total 584.72
  • x. z = 0.9945, total 582.32
  • y. z = 0, total 574
  • z. z = −0.9945, total 565.68

Why: The bigger the area to the left, the further right the percentile, so the z-values rise with the percentile. The median, the 50th percentile, has z equal to zero and sits at the centre of the sums, 574. Percentiles below the 50th have negative z-values and sit below the centre.

\[ k = 574 + (z)(8.37) \]

Two facts make this quick. First, the median of a normal total is its centre, so the 50th percentile is n times μX with no arithmetic at all. Second, the curve is symmetric, so the 16th and 84th percentiles use the same z with opposite signs and sit the same distance either side of the centre.

The 90th percentile's total, 584.72, is 574 plus 1.2816 times 8.3666, computed exactly the same way as the others.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.7 and Try It 7.7

49. Saying what a percentile of a sum means

Concept

A percentile is only half an answer until it is a sentence. The book asks for a complete sentence, and that sentence has three parts that must all be there.

\[ \text{95\% of totals of 70 sessions are at most 587.76 minutes.} \]

Compare two sentences that sound similar. Ninety-five percent of sessions last at most 587.76 minutes is wrong: it is about single sessions, and single sessions average only 8.2 minutes. Ninety-five percent of totals of 70 sessions are at most 587.76 minutes is right.

The same care applies to probabilities. A probability of 0.0127 for Example 7.5 is a statement about the total of 80 values, and about the proportion of all possible samples whose totals are that large. It is not a statement about any single value, and not about the population.

When a sentence has the share, the random variable and the units, a reader can check it against the picture without seeing any of your working.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.7(b): interpret this value in a complete sentence

50. Order the totals of 70 sessions

Ranking

All five of these values concern the total engagement time of 70 sessions, N(574, 8.37). Put them in order from smallest to largest. Try it from the picture in your head before computing anything.

Put in order

Order these totals of 70 sessions from smallest to largest.

  1. the 16th percentile
  2. the mean of the sums
  3. the 84th percentile
  4. the 95th percentile
  5. ten hours

Why: The 16th percentile is one standard deviation below the centre, 565.68 minutes; the mean of the sums is the centre, 574; the 84th percentile is one standard deviation above, 582.32; the 95th is 1.645 above, 587.76; and ten hours is 600 minutes, more than three standard deviations above.

\[ 565.68 < 574 < 582.32 < 587.76 < 600 \]

You can place every one of these by counting standard deviations of the total from the centre: minus 1, 0, plus 1, plus 1.645, plus 3.1. That is the value of standardising: once everything is in z, the ordering is obvious.

Ten hours is the item most people misplace, because it is stated in different units. Converted, it is beyond every percentile listed, which is why the probability of reaching it was one in a thousand.

51. Converting between sums and means

Section

Part 5

52. A sum is n times a mean

Concept

The total of a sample and its average carry the same information: divide the total by n and you have the mean; multiply the mean by n and you have the total.

\[ \Sigma X = n\bar{X} \]

So every event about a sum is also an event about a mean. Divide the threshold by n and the inequality survives, since n is positive.

\[ \Sigma X > 7500 \iff \frac{\Sigma X}{80} > \frac{7500}{80} \iff \bar{X} > 93.75 \]

That gives you two routes to any answer: the theorem for sums on the total, or the theorem for means on the average. They must agree, because they describe exactly the same set of samples.

The parameters convert the same way. Divide the sum's centre by n and you get μX; divide the sum's spread by n and you get the standard error.

\[ \frac{n\mu_X}{n} = \mu_X, \qquad \frac{\sqrt{n}\,\sigma_X}{n} = \frac{\sigma_X}{\sqrt{n}} \]

For Example 7.5 that is 7,200 over 80, which is 90, and 134.16 over 80, which is 1.677. Every number on the mean's ruler is the matching number on the sum's ruler divided by 80, the centre and the spread included, so a picture drawn for one is a picture drawn for the other.

This is the link between Sections 7.1 and 7.2. They are not two theorems but one, stated once for totals and once for averages.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-371 — the theorem for sums as a companion to the theorem for means

53. Why the two z-scores are always equal

Concept

Start from the z-score for a sum and divide its numerator and denominator by n. Dividing top and bottom by the same number leaves the fraction unchanged.

\[ z = \frac{\Sigma x - n\mu_X}{\sqrt{n}\,\sigma_X} \]

\[ z = \frac{(\Sigma x - n\mu_X)/n}{(\sqrt{n}\,\sigma_X)/n} \]

Now simplify each part. On top, a total over n is a mean, and n times μX over n is μX. Underneath, root n over n is one over root n.

\[ z = \frac{\bar{x} - \mu_X}{\sigma_X/\sqrt{n}} \]

That is the z-score for a mean from Section 7.1. So the two routes do not just usually agree; they give the identical z, and therefore the identical probability, every time. The only way to get different answers is to mix the routes: a sum in the numerator with a mean's spread underneath, or the other way round.

That makes the second route a perfect check. After finding a probability for a sum, divide the threshold by n, standardise it as a mean, and see the same z appear.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 371 — the z-score for ΣX

54. One curve, three rulers

Picture it

Here is Example 7.5's distribution with three rows of labels under the axis: the total, the total divided by 80, and the z-score. Nothing about the curve or the shading changes from row to row.

Figure (svg): One curve with three rows of labels under its axis. The sum row reads 6,900 to 7,500; the mean row, each sum divided by 80, reads 86.25 to 93.75; the z row reads minus 2.24 to 2.24. The yellow tail beyond sum 7,500, which is mean 93.75, is shaded.

Dividing every sum by 80 relabels the axis without moving a single point. The shaded area, and the z-score under it, cannot change.

Read down any column. A total of 7,500 is a mean of 93.75 and a z of 2.24; a total of 7,200 is a mean of 90 and a z of 0. The middle row is simply the top row divided by 80, and the spread of that row is 134.16 over 80, which is 1.677, the standard error.

\[ \frac{134.16}{80} = 1.677 = \frac{15}{\sqrt{80}} \]

So choosing between the sum route and the mean route is choosing which ruler to read. The area, which is the probability, belongs to the curve, not to the ruler.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 371-372 — Example 7.5, Figure 7.3

55. Example 7.5(a) again, by the mean route

Worked example

Re-answer Example 7.5(a), the probability that the total of 80 values exceeds 7,500, as a question about the sample mean. Population mean 90, standard deviation 15.

Convert the event

Why: divide the threshold by n equal to 80.

\[ P(\Sigma X > 7500) = P\left(\bar{X} > \frac{7500}{80}\right) = P(\bar{X} > 93.75) \]

Now the question is about the average of the 80 values, so it uses the theorem for means from Section 7.1: the centre is the population mean itself, and the spread is σX divided, not multiplied, by the square root of n.

Distribution of the mean

Why: centre μX, spread σX over root n.

\[ \bar{X} \sim N\left(90,\ \frac{15}{\sqrt{80}}\right) = N(90,\ 1.677) \]

Standardise

Why: 93.75 is 3.75 above 90.

\[ z = \frac{93.75 - 90}{1.677} = \frac{3.75}{1.677} \]

Divide

Why: the same z as the sum route.

\[ z \approx 2.236 \]

Area to the right

Why: one minus the area to the left.

\[ P(Z > 2.236) = 0.0127 \]

In context, the same sentence as before: about 1.3 percent of samples of 80 have a total over 7,500, which is to say a mean over 93.75. Two routes, one set of samples, one probability.

Use whichever route the question's numbers suit. If the question speaks of totals, the sum route keeps the units natural; if you already have the standard error from a previous part, the mean route saves a step.

Check against the sum route

Why: Example 7.5(a) gave z = 300/134.16; the two z-values must agree to rounding.

\[ \frac{300}{134.16} = 2.2361 = \frac{3.75}{1.677} \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 371-372 — Example 7.5(a)

56. Try It 7.4 re-asked as a total: 34 cans of cola

Worked example

Cans are labelled 16 ounces; they are filled with mean 16.00 ounces and standard deviation 0.143 ounces. Section 7.1 asked for the probability that the average of 34 cans exceeds 16.01 ounces. Ask it instead as a total, then check by the mean route.

Convert the event to a total

Why: 34 cans averaging over 16.01.

\[ \bar{X} > 16.01 \iff \Sigma X > (34)(16.01) = 544.34 \]

The event now concerns the total fill of 34 cans, so the theorem for sums takes over. The standard deviation per can is tiny, 0.143 ounces, and the total's spread will be only about 5.8 times that, even though the total is 34 times the size.

Centre of the sum

Why: 34 cans of 16 ounces.

\[ \mu_{\Sigma X} = (34)(16) = 544 \]

Spread of the sum

Why: root 34 is about 5.831.

\[ \sigma_{\Sigma X} = (\sqrt{34})(0.143) \approx 0.8338 \]

Standardise

Why: 0.34 ounces above the centre.

\[ z = \frac{544.34 - 544}{0.8338} = \frac{0.34}{0.8338} \approx 0.408 \]

Area to the right

Why: a sizeable area, since z is small.

\[ P(\Sigma X > 544.34) = 0.3417 \]

Figure (svg): The curve for the total fill of 34 cans, centred at 544 ounces with standard deviation 0.834. The area to the right of 544.34 ounces is shaded yellow and labelled 0.3417. Rows under the axis give the matching mean fill per can and the z-score, 0.41 at the line.

A total 0.34 ounces over the label across 34 cans is only 0.41 standard deviations of the total: it happens about a third of the time.

In context: about 34 percent of cases of 34 cans hold more than 544.34 ounces between them. The question in Section 7.1 asked whether that suggests cans are overfilled. It does not: a total that high happens about one time in three even when the mean fill is exactly 16 ounces.

The conclusion rests on the probability, not on the fact that 16.01 is above 16. A difference counts as evidence only when it would be rare if nothing were going on, and here it is common.

Check by the mean route

Why: the standard error is 0.143 over root 34, and the z must match 0.408.

\[ \frac{16.01 - 16}{0.143/\sqrt{34}} = \frac{0.01}{0.02452} \approx 0.408 \]

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.1, p. 370 — Try It 7.4

57. Translate each sum question into a mean question

Comparison

Each row is a question about a sum from this deck. Fill in the matching threshold for the sample mean, and the z-score, which the two routes share. Round to two decimal places.

Comparison matrix

sum questionnmean thresholdz (both routes)
ΣX > 7,5008093.752.24
ΣX > 1,5005030−1.89
ΣX < 1,80050360.94
ΣX ≥ 600 min708.573.11
ΣX < 200752.67-2.5

The mean thresholds are just divisions: 7,500 over 80 is 93.75, 1,500 over 50 is 30, 600 over 70 is about 8.57. The z column never changes when you switch routes; for the 1,800 row it is 100 over 106.07, or equally 2 over 2.121, about 0.94, and for the stress scores it is 25 below over a spread of 10.

If a z you compute by one route differs from the other, you have mixed a sum with a mean's spread, or the reverse. The column is there to catch that.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 371-373 — Examples 7.5-7.7

58. A percentile of the sum, read as a mean

Fill the middle

Example 7.6(c) found that the 80th percentile of the sum of 50 ages is about 1,789.27 years. Convert it to the 80th percentile of the mean age, and confirm it by the mean route. The mean's spread is 15 over root 50, about 2.1213.

Fill in the blanks

\frac35.7935.79 = ___ \qquad 34 + (0.8416)(2.1213) = ___

Why: Dividing the sum's percentile by 50 gives 35.79 years. The mean route starts at 34 and walks 0.8416 standard errors of 2.1213 up, which is 1.785, landing on the same 35.79. A percentile converts between the two scales just as a threshold does.

Both blanks come out the same, and that is the check. Eighty percent of samples of 50 users have a total age of at most 1,789.3 years, which is the same as saying they have a mean age of at most 35.79 years.

If your second blank came out far from the first, you probably walked 0.8416 of the population's 15 years, which gives 46.6. The mean's percentile uses the mean's spread.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, p. 373 — Example 7.6(c)

59. Trap: a total standardised with the mean's spread

Trap

The trap

Example 7.5(a) again. A first attempt remembers σ over root n from Section 7.1 and uses it with the total:

\[ \sigma = \frac{15}{\sqrt{80}} \approx 1.677 \]

\[ z = \frac{7500 - 7200}{1.677} \]

\[ z \approx 178.9 \]

Wrong. A total in the numerator, a mean's spread underneath.

The fix

Both parts of the fraction must describe the same random variable. Either stay with the sum:

\[ z = \frac{7500 - 7200}{134.16} \approx 2.24 \]

or convert the whole question to the mean:

\[ z = \frac{93.75 - 90}{1.677} \approx 2.24 \]

A z of 179 would make the probability zero to hundreds of decimal places, when the true answer is about one in eighty. As with the unit trap, an absurd z is a message: a numerator and a denominator from different random variables.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 371-372 — Example 7.5(a)

60. Putting it together

Section

Part 6

61. Pattern: a total is a normal question with one new line

Pattern

Every question in this section ran through the same six boxes. Only the second is new; the rest is Chapter 6.

Figure (svg): Six boxes joined by arrows in a U shape: total or average; the two parameters, centre n mu and spread root n times sigma; draw and shade; standardise with the z-score for a sum; area or percentile; say it in context.

Only box 2 is new in this section. Every other box is the normal-curve routine you already know, applied to a total.
  1. Decide the random variable: a total, a mean, or one value. Total, altogether and combined mean a sum.
  2. Write the distribution: ΣX ~ N((n)(μX), (√n)(σX)). Compute μX and σX first if you are given a shape, such as a uniform.
  3. Draw and shade: put the value or values on the axis, in the distribution's units.
  4. Standardise: z = (Σx − nμX)/(√n σX), for a probability; or pick z from the area, for a percentile.
  5. Area or walk: an area from the table, or k = nμX + z√nσX.
  6. Sentence: the share, the total of how many of what, and the units. Then check by the mean route.

\[ \Sigma X \sim N\big(n\mu_X,\ \sqrt{n}\,\sigma_X\big) \qquad z = \frac{\Sigma x - n\mu_X}{\sqrt{n}\,\sigma_X} \]

The errors this section is known for each live in one box: σ times n or the mean's spread in box 2, hours against minutes in box 3, a sentence about single values in box 6. The mean-route check catches most of them at the end.

62. Check: the distribution of a van's load

Check

Write the distribution first; everything else follows from it.

Check your understanding

A delivery van carries 40 parcels chosen at random from a warehouse where parcel weights have mean 6 kg and standard deviation 2 kg. Which is the approximate distribution of the total weight of the 40 parcels?

  • A. N(240, 12.65) (correct)
  • B. N(240, 80)
  • C. N(240, 0.32)
  • D. N(6, 0.32)

Answer: A

Why: The centre of a total is n times the mean, 40 times 6, which is 240 kg. The spread is the square root of n times the standard deviation, the square root of 40 times 2, which is about 12.65 kg.

Why B tempts people
That multiplies the standard deviation by 40 instead of by the square root of 40, as if all forty parcels were equally heavy or light together.
Why C tempts people
That uses the standard error of the mean, 2 over the square root of 40, with the total's centre. It mixes the two theorems.
Why D tempts people
That is the distribution of the mean parcel weight, not of the total load.

The wrong options are the three mistakes this deck has kept meeting: σ times n, the mean's spread with the sum's centre, and the right theorem for the wrong question. Checking the relative spread settles it: 12.65 on 240 is about 5 percent, and one parcel's spread is a third of its mean, so dividing by the square root of 40, about 6.3, is right.

\[ \frac{12.65}{240} \approx 0.053 = \frac{2/6}{\sqrt{40}} \]

That one-line check is worth making a habit. Divide the spread you found by the centre you found, and compare with one value's standard deviation over its mean, divided by the square root of n. If the two disagree, one of the parameters is wrong.

63. Check: will the van be overloaded?

Check

Same van, same parcels. Now a probability; draw the curve in your head first.

Check your understanding

The van's rated load for parcels is 260 kg. What is the approximate probability that 40 randomly chosen parcels weigh more than 260 kg in total?

  • A. about 0.057 (correct)
  • B. about 0.401
  • C. about 0.943
  • D. essentially 0

Answer: A

Why: The total follows N(240, 12.65). 260 is 20 kg above the centre, which is 20 over 12.65, about 1.58 standard deviations of the total. The area to the right of z equal to 1.58 is about 0.057.

Why B tempts people
That comes from a spread of 80 kg, σ multiplied by n. With z only 0.25 the tail looks large, but the parcels' weights are independent and partly cancel.
Why C tempts people
That is the area to the left of 260, the probability the van is not overloaded. The question asks about more than 260, the right tail.
Why D tempts people
That comes from standardising the total with the mean's spread, 0.32 kg, giving a z over 60. A numerator and a denominator from different random variables.

\[ z = \frac{260 - 240}{12.65} \approx 1.58, \qquad P(Z > 1.58) \approx 0.057 \]

Each distractor comes from a mistake with a clear signature. The answer near 0.40 means the spread was far too wide; the answer near zero means it was far too narrow; the answer near 0.94 means the wrong side of the curve was shaded. The picture in your head, with 260 a little over one and a half standard deviations up, rules out all three.

In context: about 5.7 percent of loads, roughly one van in eighteen, would be over the rated weight. That is too often to ignore, which is exactly the kind of decision a probability for a total is used for.

64. Check: a percentile for the load

Check

Same van. Now the reverse question: an area in, a total out.

Check your understanding

What is the 90th percentile of the total weight of 40 parcels?

  • A. about 256.2 kg (correct)
  • B. about 342.5 kg
  • C. about 6.4 kg
  • D. about 223.8 kg

Answer: A

Why: The z with 0.90 to its left is 1.2816. Walk from the centre 240 by 1.2816 standard deviations of the total, 1.2816 times 12.65, about 16.2 kg, to reach 256.2 kg.

Why B tempts people
That walks 1.2816 standard deviations of 80 kg, the spread you get by multiplying σ by n instead of its square root.
Why C tempts people
That is the 90th percentile of the mean parcel weight, 6 plus 1.2816 times 0.316. Multiply it by 40 and you get the right answer, 256.2.
Why D tempts people
That walks down instead of up. It is the 10th percentile: only 10 percent of totals are below it.

\[ k = 240 + (1.2816)(12.65) \approx 256.2 \]

Start with a look-up in reverse: 0.90 to the left gives z equal to 1.2816. Then the walk: 1.2816 standard deviations of the total is about 16.2 kg above the centre of 240 kg. The two numbers from the first check, the centre and the spread, do all the work.

The distractor at 6.4 kg is a useful check rather than a wrong idea: it is the right percentile on the mean's ruler, and 40 times 6.405 is 256.2. Ninety percent of van loads of 40 parcels weigh at most about 256 kg.

65. Explain the square root

Explain it to yourself

The whole section rests on one asymmetry. Explaining it in your own words is the best test of whether it has landed.

Discussion prompt

In a few sentences, explain why the mean of a total of n independent values is n times μ while its standard deviation is only the square root of n times σ. Use the two dice to make it concrete, and say what would change if the values were not independent.

A strong answer names both halves: averages add, deviations partly cancel. It uses the variance to say exactly how much they cancel, and it names independence as the condition, which is the difference between an insurance pool that works and one hit by a single flood.

If your explanation said the spread grows more slowly, push it one step further: more slowly than what, and why the square root in particular. The answer is that variances, not standard deviations, are what add.

66. Exit ticket: a café's morning takings

Exit ticket

A café serves 120 orders each morning. The amount spent per order has mean 7.40 dollars and standard deviation 3.10 dollars, and orders are independent.

Discussion prompt

Write the distribution of the morning's total takings. Find the probability that the total is under 850 dollars, and the 95th percentile of the total. Interpret the percentile in a sentence, then check the probability by the mean route.

The marks are in the order of the flow: the random variable is a total, so the theorem for sums; the spread is the square root of 120 times 3.10, not 120 times; the probability is a left tail; the percentile walks up 1.645 standard deviations of the total; and the sentence names mornings of 120 orders, not single orders.

If your two z-scores disagreed, look at the denominators: the sum route divides by 33.96, the mean route by 0.283, and 33.96 over 120 is 0.283.

67. Recap

Recap

Add up n independent values from any population with mean μX and standard deviation σX, and as n grows the total becomes normally distributed. Its centre is n times μX; its spread is the square root of n times σX.

ideaformularead it as
distribution of a sumΣX ~ N((n)(μX), (√n)(σX))normal for large n, whatever the population's shape
why √nvariances add: nσX², then take the rootindependent deviations partly cancel
relative spreadσX / (√n μX)a total gets proportionally more predictable as n grows
probabilityz = (Σx − nμX)/(√n σX)an area under the sum's curve
percentilek = nμX + z(√n)(σX)start at the centre, walk z standard deviations of the sum
sum ↔ meanΣX = nX̄; same z by both routesdivide the threshold by n to check

The skill is deciding the random variable first, a total, a mean or one value, and then keeping the square root of n in its place: multiplying σ for a total, dividing it for a mean, absent for a single value. Draw the curve, shade the event, and end with a sentence that names the total, its size and its units. Section 7.3 practises choosing between the theorems on mixed questions.

OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums §7.2, pp. 370-373 — the whole section

Sources

  1. OpenStax Introductory Statistics 2e, §7.2 The Central Limit Theorem for Sums — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0, pp. 369-374

Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.

Book on Wyzant · Text (657) 465-8108