8.1 A Single Population Mean using the Normal Distribution

The first act of statistical inference in the course. Chapter 7 asked how far a sample mean is likely to fall from a known population mean; this section reverses the question and asks which values of an unknown population mean are consistent with an observed sample mean. The answer is not a single number but an interval, formed as the point estimate plus and minus a margin of error called the error bound, which equals a z-score chosen by the confidence level multiplied by the standard error of the mean. The confidence level and its complement alpha determine that z-score, since alpha is split equally between the two tails. Raising the confidence level or shrinking the sample widens the interval, and the error bound formula can be solved backwards to find the sample size a study needs. The interpretation is the section's real difficulty: the interval is the random thing and the parameter is fixed, so the confidence level describes the procedure across repeated samples rather than any one interval.

Subject: Statistics · 68 slides · symbolic lesson

Open the interactive version of this deck

What this lesson covers

The lesson, slide by slide

1. A Single Population Mean using the Normal Distribution

Title

Statistics · §8.1

Turning one sample mean into a range of believable values for μ, and saying exactly what the confidence means

2. By the end of this lesson you can

Objectives

Chapter 7 started from a known population mean and asked where a sample mean would land. This chapter turns that around: you have one sample mean, the population mean is unknown, and you want the values of it that the sample supports.

  1. Distinguish a point estimate from an interval estimate, and write a confidence interval as (x̄ − EBM, x̄ + EBM).
  2. Turn a confidence level into α and then into the z-score that cuts off α/2 in each tail.
  3. Construct a confidence interval for μ when σ is known, from summary numbers or from raw data, and interpret it in a sentence.
  4. Say what raising the confidence level or the sample size does to the error bound, and why.
  5. Work backwards from a published interval to its error bound and sample mean, and compute the sample size a study needs.

The arithmetic is short: one z-score, one standard error, one multiplication, one plus and one minus. The difficulty is the last sentence of each problem, saying what 95 percent confident means, and the deck keeps coming back to it.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 406-416 — the section

3. What Chapter 7 handed you

Warm-up

Everything in this lesson is built from two results you already have: the standard error of a mean, and the z-scores that hold the middle of a normal curve. Recall both with the numbers of this lesson's main example.

Discussion prompt

Statistics exam scores have population standard deviation 3 points, and you take samples of 36 exams. (a) What is the standard deviation of the sample mean? (b) Which two z-scores hold the middle 90 percent of a standard normal curve between them? (c) So, within what distance of the population mean do 90 percent of sample means fall?

Part (c) is the whole lesson in embryo. If 90 percent of sample means land within 0.82 of the population mean, then for 90 percent of samples the population mean is within 0.82 of the sample mean. The rest of the deck turns that sentence around carefully and makes it precise.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 408 — the sampling distribution of x̄ and its standard error

4. Point estimate and margin of error

Section

Part 1

5. Estimating what you cannot measure

Concept

Suppose you want the mean rent of a two-bedroom apartment in your town. You cannot survey every apartment, so you write down several listed rents and average them. That average is a point estimate: one number, used as your best guess at the population mean.

point estimate — A single statistic used to estimate a population parameter. The sample mean x̄ is the point estimate of the population mean μ; the sample standard deviation s is the point estimate of σ.

\[ \bar{x} \longrightarrow \mu, \qquad s \longrightarrow \sigma \]

The same idea works for other parameters. If you want the percentage of shots you make in basketball, count the baskets and divide by the attempts; that sample proportion is the point estimate of your true proportion, the subject of Section 8.3.

Using sample data to say something about an unknown population is inferential statistics, and this chapter is where the course starts doing it. The sample mean and the sample standard deviation are statistics: they come from the sample and change from sample to sample. The population mean and standard deviation are parameters: fixed numbers you do not know.

A point estimate is almost never exactly right. A different sample of apartments would give a different average. So the honest report is not one number but a range, with a statement of how often the method that made the range succeeds. That range is a confidence interval.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8 Introduction, pp. 405-406 — point estimates and inferential statistics

6. Turning "x̄ is near μ" into "μ is near x̄"

Picture it

The book's opening example: the mean number of songs a consumer downloads a month from Apple Music. The population mean is unknown, but suppose the population standard deviation is known to be 1, and the sample has 100 people. By the central limit theorem the sample mean has standard deviation 1 over root 100.

\[ \sigma_{\bar{x}} = \frac{\sigma}{\sqrt{n}} = \frac{1}{\sqrt{100}} = 0.1 \]

The Empirical Rule says that in about 95 percent of samples the sample mean is within two standard deviations, 2 times 0.1 which is 0.2, of the population mean. That is the shaded middle of the top curve.

Figure (svg): Top: a bell curve for the sample mean, centred at the unknown mu with standard deviation 0.1, the middle region from mu minus 0.2 to mu plus 0.2 shaded and labelled 95 percent of sample means. Below, a dashed yellow line drops from mu. A green dot for a sample mean 0.13 above mu carries a bar reaching 0.2 either side, and the bar crosses the dashed line. A red dot 0.27 above mu carries the same width of bar, which stops short of the dashed line.

Distance is symmetric. Whenever the sample mean lands within 0.2 of mu, the bar of half-width 0.2 drawn around the sample mean reaches back to mu. That happens for 95 percent of samples.

Now read it backwards. Distance does not care which way you measure it: if the sample mean is within 0.2 of mu, then mu is within 0.2 of the sample mean. So the bar of half-width 0.2 drawn around a sample mean catches mu exactly when that sample mean landed in the shaded band, which happens for about 95 percent of samples.

That is the entire idea of a confidence interval. You never find out which kind of sample you have, the green kind or the red kind. You only know how often the method produces each.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8 Introduction, p. 406 — the Apple Music example and the Empirical Rule

7. The Apple Music interval

Worked example

A sample of 100 consumers gives a mean of 2 songs downloaded per month. The population standard deviation is 1. Build the approximate 95 percent interval the book builds, using the Empirical Rule's two standard deviations.

\[ \sigma = 1, \quad n = 100, \quad \bar{x} = 2 \]

Standard error of the mean

Why: the spread of sample means, not of single consumers.

\[ \frac{\sigma}{\sqrt{n}} = \frac{1}{\sqrt{100}} = \frac{1}{10} = 0.1 \]

Two standard errors

Why: the Empirical Rule's 95 percent.

\[ (2)(0.1) = 0.2 \]

The margin is 0.2 songs. It is built from the spread of the sample mean, which is why it is so small: one consumer might download five songs or none, but the average of 100 consumers is pinned down to within a few tenths.

Lower end

Why: subtract the margin from the point estimate.

\[ 2 - 0.2 = 1.8 \]

Upper end

Why: add the same margin.

\[ 2 + 0.2 = 2.2 \]

Figure (svg): A number line of mean songs downloaded per month from 1.6 to 2.4. A blue interval bar runs from 1.8 to 2.2 with the sample mean 2 marked by a yellow dot at its centre; arrows show 0.2 on each side.

The interval is the point estimate with the same margin on each side. Nothing on this line marks mu: the whole point is that you do not know where it is.

In context: we are about 95 percent confident that the mean number of songs downloaded per month from Apple Music, over the whole population of users, is between 1.8 and 2.2. The interval is (1.8, 2.2).

Check that the centre is the point estimate

Why: a symmetric interval must have the sample mean as its midpoint, and the half-length must be the margin.

\[ \frac{1.8 + 2.2}{2} = 2 = \bar{x}, \quad \frac{2.2 - 1.8}{2} = 0.2 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8 Introduction, p. 406 — the 95 percent confidence interval (1.8, 2.2)

8. Two possibilities, and you never learn which

Concept

The book is careful about what (1.8, 2.2) does and does not claim. Once the sample is taken, there are exactly two possibilities, and the interval cannot tell you which one you are in.

The second case is not a mistake. Nobody computed anything wrong; the random sample just happened to be unusual, which random samples sometimes are. A 95 percent procedure is designed to be fooled one time in twenty.

\[ P(\text{interval catches } \mu) = 0.95 \quad \text{(before sampling)} \]

An analogy helps. Before a basketball player shoots, you can say the shot has a 95 percent chance of going in, if that is their record. After the ball has gone through or bounced off, the shot is in or out, and 95 percent describes the shooter, not that shot. The interval is the shot; the method is the shooter.

Notice the words before sampling. Before you draw the sample, the interval is a random thing: its centre is a random sample mean. After you draw it, the interval is two fixed numbers, mu is a fixed number, and the interval either contains it or does not. The 95 percent is a property of the procedure, not of the particular numbers 1.8 and 2.2.

The book puts it in one line worth memorising: the confidence interval is a random variable; it is the population parameter that is fixed.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8 Introduction, pp. 405-406 — the two possibilities; the interval is random, the parameter fixed

9. Decoding the standard form of an interval

Notation

Every confidence interval in this section, and most in the rest of the book, has the same shape: a point estimate in the middle with the same amount subtracted and added. Read each piece before you use it.

Annotate

On: \( \left(\bar{x} - \text{EBM},\ \ \bar{x} + \text{EBM}\right) \)

  • The point estimate: the sample mean, computed from your data. It sits in the exact centre of the interval.
  • The error bound for a population mean, the book's name for the margin of error. It is the same amount on both sides.
  • The lower endpoint, written first.
  • The upper endpoint. The length of the whole interval is twice the error bound.
  • An interval of plausible values for the population mean μ, which does not appear in the formula because it is unknown.

News reports write the same thing two ways: a poll might say the mean is 7 with a margin of error of 2.5, or it might give the interval from 4.5 to 9.5. These are the same statement. The book's error bound is exactly what the papers call the margin of error.

\[ \bar{x} \pm \text{EBM} \quad \Longleftrightarrow \quad (\bar{x} - \text{EBM},\ \bar{x} + \text{EBM}) \]

The book adds a note: it only covers symmetric intervals, but some intervals are not symmetric, such as a confidence interval for a standard deviation. For a mean with the normal distribution, the margin is always the same on both sides.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 407 — the form of the interval, and the note on symmetry

10. Example 8.1: an interval from its two parts

Worked example

A sample has mean 7 and the error bound for the mean is 2.5. Find the confidence interval, and interpret it at a 95 percent confidence level.

\[ \bar{x} = 7, \quad \text{EBM} = 2.5 \]

Both pieces are handed to you, so there is nothing to compute but the two endpoints. In later examples the error bound itself has to be built, which is where the confidence level enters.

Lower endpoint

Why: point estimate minus the error bound.

\[ 7 - 2.5 = 4.5 \]

Upper endpoint

Why: point estimate plus the error bound.

\[ 7 + 2.5 = 9.5 \]

Write the interval

Why: lower endpoint first.

\[ (4.5,\ 9.5) \]

The interval is 5 units long, twice the error bound. Notice that the confidence level played no part in the arithmetic: it was already built into the error bound of 2.5 that the problem handed you. In the next examples you will build that error bound yourself, and that is where the level does its work.

The book's sentence: we estimate with 95 percent confidence that the true value of the population mean is between 4.5 and 9.5. The sentence names the level, the parameter, and both endpoints; with a real context it would also name the population and the units.

Check the midpoint and half-length

Why: they must give back the two parts you started from.

\[ \frac{4.5 + 9.5}{2} = 7, \quad \frac{9.5 - 4.5}{2} = 2.5 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 407 — Example 8.1

11. Try It 8.1: assemble the interval

Fill the middle

A sample has mean 15 and the error bound for the mean is 3.2. Fill in both endpoints. Then read the check under the box, which is the habit to build for every interval you write.

Fill in the blanks

\left(15 - 3.2,\ \ 15 + 3.2\right) = \left(11.8,\ 18.2\right)

Why: Subtract the error bound for the lower end and add it for the upper end: 15 minus 3.2 is 11.8 and 15 plus 3.2 is 18.2. The interval is 6.4 long, twice the error bound.

\[ \frac{11.8 + 18.2}{2} = 15, \qquad 18.2 - 11.8 = 6.4 = 2(3.2) \]

The midpoint of the two endpoints must be the sample mean, and the length must be twice the error bound. If either fails, an endpoint is wrong. This check costs ten seconds and catches the common slip of adding the error bound twice or subtracting it from the wrong number.

The same two facts, the centre is the sample mean and the half-length is the error bound, are what Part 5 uses to take a published interval apart.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 407 — Try It 8.1

12. Which part of a confidence statement is random?

Prediction

Before any data are collected, a researcher plans to take a random sample of 100 users and build the interval from the sample mean plus and minus 0.2. Think about what could come out differently if the study were run again.

Predict first

Before the sample is drawn, which of these is a random quantity?

  • the interval's endpoints, because they are built from the sample mean
  • the population mean, because it is unknown
  • the margin of 0.2, because it depends on the sample
  • the confidence level, because it is a probability

Correct: the interval's endpoints, because they are built from the sample mean

Why: The endpoints are the sample mean minus and plus 0.2, and the sample mean changes from sample to sample, so the endpoints do too. The population mean is unknown but fixed. With sigma known, the margin depends only on sigma, n and the level, so it is the same for every sample. The level is a fixed choice made in advance.

Unknown is not the same as random. The population mean is one definite number; you just do not know it. What varies from one run of the study to the next is the sample, hence the sample mean, hence where the interval sits.

\[ \underbrace{\bar{x}}_{\text{random}} \pm \underbrace{1.96\,\frac{\sigma}{\sqrt{n}}}_{\text{fixed when } \sigma \text{ known}} \]

This is why the margin can be quoted before a poll is even run. A pollster who knows sigma, the sample size and the level knows the error bound in advance; only the centre waits for the data.

That is why the probability statement belongs to the interval: before sampling, the random interval has a 95 percent chance of landing on the fixed mu, the way a thrown ring has a chance of landing on a fixed peg.

13. Trap: a probability about μ after the fact

Trap

The trap

The Apple Music interval is (1.8, 2.2) at 95 percent confidence. A common sentence:

Wrong: there is a 95 percent probability that the population mean is between 1.8 and 2.2.

\[ P(1.8 < \mu < 2.2) = 0.95 \quad \times \]

Mu is a fixed number and so are 1.8 and 2.2. That event is either certainly true or certainly false; it has no 95 percent chance.

The fix

The 95 percent belongs to the method, measured over repeated samples.

Right: we estimate with 95 percent confidence that the mean number of songs downloaded per month is between 1.8 and 2.2.

And what it means: about 95 percent of intervals built this way, from repeated samples, contain the true mean.

The book flags the trap itself: the confidence level is often taken as the probability the computed interval contains the parameter, but it is more accurate to call it the percent of intervals that contain the parameter when repeated samples are taken.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 407 — what the confidence level means

14. The confidence level and alpha

Section

Part 2

15. The confidence level is an area in the middle

Concept

The Empirical Rule's two standard deviations gave about 95 percent. To hit a chosen level exactly, you need the z-scores that hold exactly that much area in the middle of the standard normal curve. Take the book's 90 percent.

Figure (svg): A standard normal curve with the middle 0.90 of its area shaded blue between z of minus 1.645 and 1.645, both marked by yellow lines. The two tails beyond those lines are shaded red and each labelled alpha over two equals 0.05.

The confidence level is the middle area, and alpha is what is left, split evenly. The cut points are the z-scores that leave exactly 0.05 in each tail.

Holding the central 90 percent leaves 10 percent outside, and by symmetry that 10 percent is split evenly, 5 percent in each tail. The cut points are the z-scores with 0.05 beyond them on each side, which are minus 1.645 and 1.645.

alpha (α) — The probability that the interval does not contain the population parameter: the area left in the two tails. Alpha and the confidence level add to one.

\[ \alpha + \text{CL} = 1 \quad \Rightarrow \quad \alpha = 1 - 0.90 = 0.10 \]

\[ \text{each tail} = \frac{\alpha}{2} = \frac{0.10}{2} = 0.05 \]

So to be 90 percent confident, the interval reaches 1.645 standard errors either side of the sample mean. Most people building an interval choose 90 percent or higher, because they want to be reasonably sure of their conclusion.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 407-408 — alpha, CL, and Figure 8.2

16. Why alpha is split evenly between the tails

Concept

Nothing forces the 10 percent of a 90 percent interval to be split five and five. You could leave 2 percent in the left tail and 8 percent in the right, and the middle would still hold 90 percent. So why does every interval in this section split alpha evenly?

Figure (svg): Two standard normal curves, one above the other, each with 0.90 of its area shaded. Top, in blue: the symmetric region from minus 1.645 to 1.645, 3.290 wide. Bottom, in orange: a lopsided region from minus 2.054 to 1.405, leaving 0.02 in the left tail and 0.08 in the right, 3.459 wide. A bar under each axis shows the width.

Both regions hold exactly 90 percent. The symmetric one is shorter, because it keeps the tallest part of the curve, so it gives the narrowest interval for the same confidence.

Compare the widths. The even split runs from minus 1.645 to 1.645, which is 3.290 standard errors long. The lopsided split runs from minus 2.054 to 1.405, which is 3.459 standard errors, about 5 percent longer, for exactly the same confidence.

\[ 1.645 - (-1.645) = 3.290 \quad < \quad 1.405 - (-2.054) = 3.459 \]

The reason is the shape of the curve. Near the centre the normal curve is tall, so a short stretch holds a lot of area; out in a tail it is low, so you need a long stretch for the same area. The even split takes its area from the tallest part on both sides, and any lopsided split trades some of that for thinner tail, which costs width.

So for a symmetric curve, the even split gives the shortest interval at a given level, and the interval comes out symmetric about the sample mean. That is where the book's symmetric form, the sample mean plus and minus the same error bound, really comes from.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 409 — alpha is split equally between the two tails

17. Decoding z sub alpha over two

Notation

The book names the cut point with a subscript that tells you the area to its right. It looks fussy, but every piece of the symbol is an instruction.

Annotate

On: \( z_{\alpha/2}, \quad \text{e.g. } z_{0.025} = \text{invNorm}(0.975,\ 0,\ 1) = 1.96 \)

  • One minus the confidence level: the total area left in the two tails. For 95 percent, alpha is 0.05.
  • Half of alpha: the area in one tail. The subscript gives the area to the RIGHT of the cut point.
  • The z-score with 0.025 to its right, so 0.975 to its left.
  • The calculator wants the area to the LEFT, so you enter 1 minus alpha over 2. The 0 and 1 say: standard normal.
  • The answer, a positive number: the cut point on the right. The left cut point is its negative.

The single trap in this notation is the direction of the area. The subscript counts the right tail; the calculator and most tables count from the left. So the number you enter is one minus alpha over two, never alpha over two itself.

\[ \text{area to the left} = 1 - \frac{\alpha}{2} \]

If you enter alpha over two by mistake you get the left cut point, a negative z-score, and an error bound that is negative. That is the first clue something went wrong: an error bound is a distance, so it is never negative.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 409 — z sub alpha over two and the invNorm note

18. The z-score for a 90 percent level

Worked example

Example 8.2 will need the z-score for a 90 percent confidence level. Find it in the four moves you will use for every level, and check it against the Empirical Rule.

\[ \text{CL} = 0.90 \]

Find alpha

Why: alpha plus the confidence level is one.

\[ \alpha = 1 - 0.90 = 0.10 \]

Split alpha between the tails

Why: the interval is symmetric, so each tail gets half.

\[ \frac{\alpha}{2} = 0.05 \]

Convert to an area to the left

Why: the right tail holds 0.05, so 0.95 lies to the left of the cut.

\[ 1 - 0.05 = 0.95 \]

Now go to the standard normal. You want the z-score whose left area is 0.95. A table lists 0.9495 at 1.64 and 0.9505 at 1.65, so the value is halfway between; a calculator gives it directly.

Read the z-score

Why: invNorm(0.95, 0, 1).

\[ z_{0.05} = 1.645 \]

Keep this number. It is the 90 percent multiplier for every interval in this section, whatever the context, because the standard normal curve does not care what you are measuring.

Check against the Empirical Rule

Why: 90 percent sits between 68 percent (one standard deviation) and 95 percent (about two), so its z must sit between 1 and 2, and much nearer 2.

\[ P(-1.645 < Z < 1.645) = 0.95 - 0.05 = 0.90 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 408-410 — Figure 8.2 and Example 8.2

19. Finding z from a printed table

Concept

The book allows a calculator, a computer, or a standard normal table. A table runs the other way from what you need: it takes a z-score and gives the area to its left. So you look for your area in the body of the table and read the z from the margins.

zarea to the leftused for
1.640.9495
1.650.950590%: want 0.9500, halfway, so 1.645
1.960.975095%: exactly 0.9750, so 1.96
2.320.9898
2.330.990198%: want 0.9900, nearer 2.33, so 2.326
2.570.9949
2.580.995199%: want 0.9950, halfway, so 2.575 or 2.576

Two of the four common levels land between table rows. For 90 percent the area 0.95 sits exactly halfway between 0.9495 and 0.9505, so the z is halfway between 1.64 and 1.65. That is why the book writes 1.645, with three decimals, while 1.96 needs only two.

\[ \text{area } 0.9500 \text{ between } 0.9495 \text{ and } 0.9505 \Rightarrow z = 1.645 \]

Whatever tool you use, the input is the same: the area to the left of the upper cut, one minus alpha over two. The table just makes you do the lookup by eye. Expect the third decimal place to differ slightly between a table and a calculator; it never moves an interval's endpoints by more than a rounding step.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 409 — using a calculator, computer or a standard normal probability table

20. From 99 percent to its z-score

Fill the middle

Run the same four moves for a 99 percent level, the highest of the four levels you will meet most often. Fill in alpha, the tail area, and the left area you would give the calculator.

Fill in the blanks

\alpha = 1 - 0.99 = 0.01, \quad \frac0.0050.995 = ___, \quad z____ = \text___(___) = 2.576

Why: Alpha is 1 minus 0.99, which is 0.01. Each tail gets half, 0.005. The calculator wants the area to the left of the right-hand cut, which is 1 minus 0.005, so 0.995, and the z-score with that much to its left is 2.576.

The z-score has grown from 1.645 at 90 percent to 2.576 at 99 percent. Squeezing the tails from 5 percent each down to half a percent each pushes the cuts a long way out, because the normal curve is thin out there and you have to travel far to shed a little more area.

\[ \text{CL } 0.90 \to 1.645, \quad 0.95 \to 1.96, \quad 0.98 \to 2.326, \quad 0.99 \to 2.576 \]

Those four pairs are worth knowing by sight. They also give a fast check on any z you compute: a higher level must give a larger z.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 409 — finding the z-score for the stated confidence level

21. Four confidence levels on one curve

Picture it

Put the four common levels on a single standard normal curve and the trade-off of the whole section is visible before you compute any interval.

Figure (svg): A standard normal curve with four pairs of dashed vertical lines, symmetric about zero: green at plus and minus 1.645 for 90 percent, blue at 1.96 for 95 percent, orange at 2.326 for 98 percent and red at 2.576 for 99 percent. The labels sit above the curve.

Higher confidence pushes the cut points further into the tails. Going from 90 to 99 percent moves the cut from 1.645 to 2.576, more than half again as far.

Each pair of dashed lines encloses one level's middle area. The 90 percent pair is closest to the centre; the 99 percent pair is furthest out. The level grows by only 9 percentage points from the inner pair to the outer pair, but the z-score grows by more than half.

\[ \frac{2.576}{1.645} = 1.57 \]

Since the error bound is z times the standard error, the 99 percent interval from any sample is 1.57 times as wide as the 90 percent interval from the same sample. More confidence is never free: you pay for it in width.

The book's Figure 8.5 makes the same comparison for 90 and 95 percent. Part 4 builds both of those intervals for the exam scores and lays them side by side.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 409, 414 — the z-scores, and Figure 8.5

22. Estimate the z-score for 93 percent

Estimation

Try It 8.3 asks for a 93 percent interval, an unusual level that is not on the list you have memorised. Before computing, place its z-score among the ones you know.

Predict first

Where does the z-score for a 93 percent confidence level fall?

  • between 1.645 and 1.96
  • between 1.96 and 2.326
  • below 1.645
  • exactly halfway between 1.645 and 1.96, at about 1.80

Correct: between 1.645 and 1.96

Why: Ninety-three percent lies between 90 and 95 percent, and higher levels have larger z-scores, so its z lies between 1.645 and 1.96. It is not exactly halfway, because the curve is not straight: the computed value is 1.812, a little past the midpoint 1.80.

\[ \alpha = 0.07, \quad \frac{\alpha}{2} = 0.035, \quad z_{0.035} = \text{invNorm}(0.965) = 1.812 \]

Bracketing is a good habit because it catches the commonest z mistake, forgetting to halve alpha. If you had entered 0.93 into invNorm you would get 1.476, which is below 1.645 and so cannot belong to a level above 90 percent.

The halfway guess is close but not right. Between two cut points the area under the normal curve is not spread evenly, so equal steps in confidence do not give equal steps in z. Use the calculator for the exact value and the bracket to check it.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 412 — Try It 8.3 uses a 93 percent level

23. Four calculator entries for a 95 percent z

Error analysis

Four students find the z-score for a 95 percent interval with the calculator. Only one entry is right. Find the fault in each of the others; each of these slips turns up every term.

Annotate

On: \( z = 1.645 \quad z = -1.645 \quad z = -1.96 \quad z = 1.96 \)

  • Gives the confidence level to invNorm without splitting alpha. That puts all 5 percent in one tail: it is the 90 percent z, so the interval comes out too narrow.
  • Enters alpha itself as a left area. Both mistakes at once: alpha not halved, and the left tail used. The negative sign gives it away.
  • Right tail area, wrong side: 0.025 to the LEFT is the lower cut. The size is right, but used as a negative error bound it swaps the endpoints.
  • Correct: 0.025 in the right tail means 0.975 to the left, so z sub 0.025 is 1.96.

The first error is the dangerous one, because 1.645 is a perfectly real z-score, just for a different level. Nothing about it looks wrong. The defence is the ordering: a 95 percent interval must be wider than a 90 percent one, so its z must be bigger than 1.645.

The two negative answers announce themselves. The book's z sub alpha over two is always the right-hand cut point, a positive number, because it is going to be multiplied by a standard error to make a distance.

\[ \text{invNorm}\left(1 - \tfrac{\alpha}{2}\right) > 0 \quad \text{always} \]

When you check your own work, ask two questions about the number you got. Is it positive? Is it bigger than the z for any lower level? A 95 percent z that fails either test came from the wrong area.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 409 — the note: remember to use the area to the left

24. Match each level to its z-score

Matching

Six confidence levels and six z-scores. Match them, using the order rule as you go: a higher level always pairs with a larger z. Two of these levels are new, and you can place them by bracketing.

Match the pairs

  • a. 80%
  • b. 90%
  • c. 93%
  • d. 95%
  • e. 98%
  • f. 99%
  • p. 1.282
  • q. 1.645
  • r. 1.812
  • s. 1.960
  • t. 2.326
  • u. 2.576

Why: Each z is invNorm of 1 minus half of alpha: 0.90 gives 1.282 for 80 percent, 0.95 gives 1.645, 0.965 gives 1.812, 0.975 gives 1.960, 0.99 gives 2.326 and 0.995 gives 2.576. Because z grows with the level, sorting both lists in increasing order pairs them.

The ordering does most of the work: once both lists are sorted, the pairs line up. That is the check worth doing every time you look up a z.

\[ \text{CL } 0.80: \quad \alpha/2 = 0.10, \quad \text{invNorm}(0.90) = 1.282 \]

If you matched by memory and got stuck on the two unfamiliar levels, bracket them. Eighty percent is below 90, so its z is below 1.645, and only one z on the list is. Ninety-three percent sits between 90 and 95, and only one z sits between 1.645 and 1.96.

The 80 percent value, 1.282, is the same number Chapter 7 used for a 90th percentile. That is no coincidence: an 80 percent middle leaves 10 percent above it, so its upper cut is the 90th percentile.

25. Constructing the interval

Section

Part 3

26. Where the formula comes from

Intuition

The interval is not a recipe to memorise; it falls out of the sampling distribution in four lines. Start from the fact that the sample mean is normal with mean mu and standard deviation sigma over root n, so its z-score is standard normal.

\[ P\left(-z_{\alpha/2} < \frac{\bar{X} - \mu}{\sigma/\sqrt{n}} < z_{\alpha/2}\right) = 1 - \alpha \]

Multiply all three parts by the standard error. The middle becomes the distance between the sample mean and mu.

\[ P\left(-z_{\alpha/2}\tfrac{\sigma}{\sqrt{n}} < \bar{X} - \mu < z_{\alpha/2}\tfrac{\sigma}{\sqrt{n}}\right) = 1 - \alpha \]

Call that distance the error bound. Then rearrange the inequality to put mu alone in the middle: subtract the sample mean, multiply by minus one, and the inequality signs flip.

\[ \text{EBM} = z_{\alpha/2}\,\frac{\sigma}{\sqrt{n}} \]

\[ P\left(\bar{X} - \text{EBM} < \mu < \bar{X} + \text{EBM}\right) = 1 - \alpha \]

Read the last line slowly. The random thing is the sample mean, so the random things are the endpoints. Before sampling, the interval from the sample mean minus EBM to plus EBM catches mu with probability equal to the confidence level. That is exactly the claim, and exactly why the claim is about the procedure.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 408-410 — the confidence interval formula

27. Why the standard error, and not σ

Concept

The book insists that the standard deviation in the error bound must be the one appropriate to the parameter being estimated. You are estimating a mean, so you need the spread of sample means, sigma over root n, which is called the standard error of the mean.

Figure (svg): Two normal curves centred at mu on an axis of exam scores from mu minus 10 to mu plus 10. A low wide orange curve is one exam score, standard deviation 3. A tall narrow blue curve is the mean of 36 scores, standard deviation 0.5, with its middle 90 percent shaded, about 0.82 either side of mu.

The interval estimates a mean, so its width comes from the narrow curve. Using the wide curve would build an interval six times too wide.

For the exam scores, one exam varies with standard deviation 3 points, but the mean of 36 exams varies with standard deviation 0.5. The interval is about where mu is, and sample means are what tell you, so their spread sets the width.

\[ \frac{\sigma}{\sqrt{n}} = \frac{3}{\sqrt{36}} = 0.5 \quad \text{not} \quad \sigma = 3 \]

The name standard error exists to keep these apart: the population standard deviation sigma describes individuals; the standard error describes the mean. Using sigma would give an interval six times too wide here, a range covering most of the individual exam scores rather than the plausible values of their mean.

In summary, as the book puts it: by the central limit theorem, the sample mean is normal with mean mu and standard deviation sigma over root n, and when sigma is known, a normal distribution gives the error bound.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 408 — the standard error of the mean

28. When this interval is allowed

Concept

The formula in this section rests on one claim: the sample mean is normally distributed with standard deviation sigma over root n. Three conditions make that claim true, and an interval built without them can be badly wrong however carefully the arithmetic is done.

conditionwhy it mattersin this section's examples
a random samplethe theory of sampling distributions assumes it; a biased sample shifts x̄ away from μevery example says random
σ knownthe error bound uses σ itself; with σ unknown, s replaces it and the curve changesσ = 3, 6, 0.337, 15
population normal, or n largethen x̄ is normal, exactly or by the central limit theoremexam scores normal; SAR n = 30

The first condition is the one no formula can fix. If the sample is a convenience sample, such as the people who answer an online poll, the interval is centred in the wrong place and its 95 percent means nothing.

The second is the one this section takes for granted and the next section drops. Sigma is almost never known in real life; the book assumes it so that the normal curve applies. When you have to estimate sigma by s, Section 8.2 replaces z with a Student's t value.

\[ \text{this section: } \bar{x} \pm z_{\alpha/2}\frac{\sigma}{\sqrt{n}} \qquad \text{8.2: } \bar{x} \pm t_{\alpha/2}\frac{s}{\sqrt{n}} \]

The third is the central limit theorem's condition from Chapter 7. A normal population makes the sample mean normal for any n; otherwise you need a reasonably large sample, often taken as 30 or more, and more for strongly skewed data.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 407-408 — the interval is based on the central limit theorem, with σ known

29. Example 8.2: a 90 percent interval for exam scores

Worked example

Statistics exam scores are normal with unknown mean and population standard deviation 3 points. A random sample of 36 scores has mean 68. Find a 90 percent confidence interval for the population mean exam score.

\[ \sigma = 3, \quad n = 36, \quad \bar{x} = 68, \quad \text{CL} = 0.90 \]

The point estimate

Why: the sample mean is the centre of the interval.

\[ \bar{x} = 68 \]

The z-score for 90 percent

Why: alpha is 0.10, so 0.05 in each tail; invNorm(0.95, 0, 1).

\[ z_{0.05} = 1.645 \]

The standard error

Why: the spread of the mean of 36 scores.

\[ \frac{\sigma}{\sqrt{n}} = \frac{3}{\sqrt{36}} = \frac{3}{6} = 0.5 \]

The error bound

Why: z standard errors.

\[ \text{EBM} = (1.645)(0.5) = 0.8225 \]

With the exact z, 1.6449, the error bound is 0.8224; both round the endpoints the same way. The margin is less than one point, because the mean of 36 exams is far steadier than a single exam.

Lower endpoint

Why: subtract.

\[ 68 - 0.8225 = 67.1775 \approx 67.18 \]

Upper endpoint

Why: add.

\[ 68 + 0.8225 = 68.8225 \approx 68.82 \]

Figure (svg): A bell curve centred at the sample mean 68 with standard error 0.5, its middle 0.90 shaded between 67.18 and 68.82, both marked with yellow lines. Below the axis, a blue interval bar with a yellow centre dot at 68 runs from 67.18 to 68.82, labelled x-bar plus or minus 0.82.

The interval is 1.645 standard errors either side of 68. A standard error here is half a point, so the margin is about 0.82 of a point.

In context: we estimate with 90 percent confidence that the true population mean exam score for all statistics students is between 67.18 and 68.82. The calculator's ZInterval, with sigma 3, mean 68, n of 36 and level 0.90, prints (67.178, 68.822).

Check by standardising an endpoint

Why: the upper end should sit exactly 1.645 standard errors above the sample mean.

\[ \frac{68.8225 - 68}{0.5} = 1.645 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 409-410 — Example 8.2, Figure 8.3

30. A hundred intervals from one population

Picture it

The book explains Example 8.2's 90 percent like this: if we constructed 100 of these intervals, we would expect 90 of them to contain the true mean. Here is that experiment, run in the deck's code. Pretend the true mean is 67.5, draw 100 samples of 36 exams, and build the 90 percent interval from each.

Figure (svg): A hundred vertical bars side by side, each a 90 percent confidence interval from a different simulated sample of 36 exam scores drawn from a population with mean 67.5 and standard deviation 3. A dashed yellow horizontal line marks the true mean 67.5. 86 bars, in blue, cross the line; 14, in red, sit wholly above or below it. Every bar has the same length, 1.64 points.

The line stays put and the bars move. The 90 percent describes the method: about nine in ten bars it produces cross the line, and nothing on any one bar tells you whether it is a hit.

In this run, 86 of the 100 intervals contain 67.5 and 14 miss. Every interval has the same length, 2 times 0.8224, because sigma, n and the level never changed. Only the centres moved, following the sample means up and down.

Now imagine you are a researcher holding one of these bars, and you cannot see the yellow line. Your bar might be blue or red. All you can say is that your method produces blue bars about 90 percent of the time. That is exactly what we estimate with 90 percent confidence means.

\[ \text{about } 90 \text{ of } 100 \text{ intervals catch } \mu \]

The count is not exactly 90 in a given run, for the same reason that 100 coin tosses do not give exactly 50 heads. Over very many runs the proportion settles at 90 percent.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 411 — explanation of the 90 percent confidence level

31. Try It 8.2: pizza delivery times

Worked example

Pizza delivery times are normal with unknown mean and population standard deviation 6 minutes. A random sample of 28 restaurants has mean delivery time 36 minutes. Find a 90 percent confidence interval for the population mean delivery time.

\[ \sigma = 6, \quad n = 28, \quad \bar{x} = 36, \quad \text{CL} = 0.90 \]

The z-score

Why: the same 90 percent value as Example 8.2.

\[ z_{0.05} = 1.645 \]

The standard error

Why: 28 is not a perfect square, so keep four decimals.

\[ \frac{6}{\sqrt{28}} = \frac{6}{5.2915} = 1.1339 \]

The error bound

Why: z standard errors.

\[ \text{EBM} = (1.645)(1.1339) = 1.8653 \]

The steps are identical to Example 8.2, only the numbers differ. The error bound is almost two minutes this time: a bigger sigma and a smaller sample both make the mean less precise than the exam mean was.

Lower endpoint

Why: subtract.

\[ 36 - 1.8653 = 34.13 \]

Upper endpoint

Why: add.

\[ 36 + 1.8653 = 37.87 \]

In context: we estimate with 90 percent confidence that the mean delivery time for all pizza delivery restaurants is between 34.13 and 37.87 minutes. With the exact z, 1.6449, the endpoints are 34.135 and 37.865, the same to two decimals.

Check the midpoint and the z

Why: the midpoint must be 36, and the half-width over the standard error must be the z.

\[ \frac{34.13 + 37.87}{2} = 36, \quad \frac{1.8653}{1.1339} = 1.645 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 411 — Try It 8.2

32. Example 8.3: before you compute

Step zero

Example 8.3 is about cell phone radiation. The Specific Absorption Rate, SAR, measures the radio-frequency energy a user's body absorbs from a handset; to be sold in the United States a phone must have SAR at most 1.6 watts per kilogram. The book lists the highest SAR of 30 randomly selected models, and asks for a 98 percent interval for the true mean SAR, with sigma taken as 0.337.

Discussion prompt

Before any arithmetic: what is the parameter being estimated, in words? What is the point estimate, and does the book give it to you? Which three numbers go into the error bound, and what is each one here?

Taking stock first stops two mistakes this example invites. The sample standard deviation of these 30 values is 0.403, and it is tempting to use it; but the book states sigma is 0.337, and in this section sigma is the known population value. The other mistake is using the FCC limit, 1.6, somewhere in the calculation. It is context, not data.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 411 — Example 8.3, Table 8.1

33. Example 8.3(a): the sample mean of 30 phones

Worked example

Start with the point estimate. The 30 SAR values from Table 8.1, in watts per kilogram, are added and divided by 30. Group them in rows of ten so the addition can be checked.

\[ 1.11 + 1.36 + 0.74 + 1.48 + 1.34 + 0.5 + 1.43 + 1.18 + 0.4 + 1.3 = 10.84 \]

\[ 1.3 + 0.867 + 1.09 + 1.26 + 0.68 + 0.455 + 1.29 + 0.51 + 1.41 + 0.36 = 9.222 \]

\[ 1.13 + 0.82 + 0.52 + 0.3 + 0.78 + 1.6 + 1.48 + 1.25 + 1.39 + 1.38 = 10.65 \]

Total

Why: add the three row sums.

\[ 10.84 + 9.222 + 10.65 = 30.712 \]

Divide by n

Why: thirty phones.

\[ \bar{x} = \frac{30.712}{30} = 1.0237 \]

Figure (svg): A dot plot of the 30 phone SAR values from 0.3 to 1.6 watts per kilogram, stacked in bins of 0.05; most dots lie between 1.1 and 1.5 with a scattered left tail down to 0.3. A dashed red line marks the FCC limit 1.6. Above the dots, a yellow interval bar runs from 0.881 to 1.167 with its centre at the sample mean 1.024.

The interval is far narrower than the spread of the phones themselves. It estimates the mean SAR of all phones, not the SAR of any one phone.

The book rounds to 1.024. The picture shows why a mean is worth estimating carefully here: the phones themselves range from 0.3 to 1.6, the left tail is long, and individual values tell you little about the centre. The interval drawn above the dots is the one part (b) builds.

The FCC line at 1.6 is useful context but plays no part in the arithmetic. One phone in the sample sits exactly at the limit.

Check the mean against the dot plot

Why: the mean should sit a little left of the main cluster, pulled down by the low values.

\[ 1.0237 < \text{median} = 1.255 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 411-412 — Example 8.3, Table 8.1

34. Example 8.3(b): the 98 percent interval for mean SAR

Worked example

Now the interval: 98 percent confidence, sigma 0.337, n of 30, and the sample mean 1.024 from part (a).

The z-score for 98 percent

Why: alpha 0.02, so 0.01 per tail and 0.99 to the left: invNorm(0.99, 0, 1).

\[ z_{0.01} = 2.326 \]

The standard error

Why: sigma over root 30.

\[ \frac{0.337}{\sqrt{30}} = \frac{0.337}{5.4772} = 0.06153 \]

The error bound

Why: z standard errors.

\[ \text{EBM} = (2.326)(0.06153) = 0.1431 \]

A 98 percent level is demanding, so the multiplier is large, 2.326. But 30 phones and a small sigma keep the standard error to about six hundredths, so the margin is still only 0.14 watts per kilogram.

Lower endpoint

Why: subtract from the point estimate.

\[ 1.024 - 0.1431 = 0.8809 \]

Upper endpoint

Why: add to the point estimate.

\[ 1.024 + 0.1431 = 1.1671 \]

A misprint to watch for: the book's solution writes the upper line as x-bar minus EBM, 1.024 minus 0.1431, equals 1.1671. The arithmetic shows it must be plus. Using the unrounded mean 1.0237 gives (0.8806, 1.1669); both round to the calculator's (0.881, 1.167).

In context: we estimate with 98 percent confidence that the true mean SAR for the population of cell phones in the United States is between 0.8809 and 1.1671 watts per kilogram.

Check the width

Why: the interval's length must be twice the error bound.

\[ 1.1671 - 0.8809 = 0.2862 = 2(0.1431) \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 412 — Example 8.3, Figure 8.4

35. Try It 8.3: a second sample of 20 phones at 93 percent

Worked example

Table 8.2 gives the highest SAR of a different random sample of 20 models. Use it to build a 93 percent confidence interval for the true mean SAR, again with sigma 0.337.

\[ 1.48 + 1.53 + 0.8 + 0.68 + 1.15 + 1.4 + 1.36 + 1.24 + 0.77 + 0.57 = 10.98 \]

\[ 0.462 + 0.2 + 1.36 + 0.51 + 1.39 + 0.3 + 1.3 + 0.73 + 0.7 + 0.869 = 7.821 \]

Sample mean

Why: both rows, over 20.

\[ \bar{x} = \frac{10.98 + 7.821}{20} = \frac{18.801}{20} = 0.9400 \]

The z-score for 93 percent

Why: alpha 0.07, so 0.035 per tail, 0.965 to the left.

\[ z_{0.035} = \text{invNorm}(0.965) = 1.812 \]

Standard error and error bound

Why: sigma over root 20, times z.

\[ \frac{0.337}{\sqrt{20}} = 0.07536, \quad \text{EBM} = 1.812(0.07536) = 0.1365 \]

The interval

Why: subtract and add.

\[ (0.9400 - 0.1365,\ 0.9400 + 0.1365) = (0.8035,\ 1.0765) \]

In context: we estimate with 93 percent confidence that the true mean SAR of certified cell phones is between 0.8035 and 1.0765 watts per kilogram.

Figure (svg): Two dot plots on the same SAR axis from 0.1 to 1.7. The top one shows the 30 phones of Example 8.3 in blue with its 98 percent interval (0.881, 1.167) drawn above. The bottom one shows the 20 phones of Try It 8.3 in orange with its 93 percent interval (0.804, 1.077). A short yellow band between them marks the stretch 0.881 to 1.077 that both intervals share.

Different phones, a different sample size and a different level give a different interval. The two overlap by a wide stretch, so they do not contradict each other.

The book's comment: the two intervals differ because they come from different samples, of different sizes, at different levels. They do not conflict; they overlap from 0.881 to 1.077, and any mean in that stretch is consistent with both.

Check the z by bracketing

Why: 93 percent lies between 90 and 95 percent, so its z must lie between theirs.

\[ 1.645 < 1.812 < 1.960 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 412-413 — Try It 8.3, Table 8.2

36. Writing the interpretation sentence

Concept

The book asks for a sentence after every interval, and gives a template. It has four slots, and each slot is something you decided while computing.

slotwhat goes in itExample 8.3
the levelthe confidence level as a percent98% confidence
the parameterthe population mean, with its population namedthe true mean SAR of all US cell phones
the endpointsboth numbers, lower firstbetween 0.8809 and 1.1671
the unitsthe units of the measurementwatts per kilogram

The template itself: we estimate with blank percent confidence that the true population mean, with context, is between blank and blank, with units. Leave out any slot and the sentence stops being an interpretation and becomes a restatement of the numbers.

The slot that goes wrong is the parameter. The interval is about the mean of the population, not about individual phones, not about the sample mean (which you know exactly, 1.024), and not about future sample means. If your sentence would make sense with the word phones in place of mean SAR of phones, rewrite it.

A second sentence explains the level, and the book writes one for Example 8.2: ninety percent of all confidence intervals constructed in this way contain the true mean statistics exam score. That sentence is about the method; the first one is about this interval.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 409, 411 — writing the interpretation, and the explanation of the level

37. Which readings of (67.18, 68.82) are right?

Sorting

Example 8.2 gave the 90 percent interval (67.18, 68.82) for the mean statistics exam score. Six people describe what it means. Sort each statement, and look for what it is describing: the population mean, individual exams, sample means, or the method.

Sort into buckets

Sort each statement about the 90 percent interval (67.18, 68.82).

a correct reading
We estimate with 90% confidence that the mean score of all statistics exams is between 67.18 and 68.82.; If we sampled 36 exams many times, about 90% of the intervals built this way would contain μ.; This interval came from a method that captures the population mean 90% of the time.
an incorrect reading
There is a 90% probability that μ is between 67.18 and 68.82.; 90% of all statistics exam scores are between 67.18 and 68.82.; 90% of samples of 36 exams have a sample mean between 67.18 and 68.82.
ok
The book's sentence, the repeated-sampling meaning, and the method statement all put the 90 percent where it belongs: on the procedure, across many samples, with the population mean as the target.
no
The probability statement treats the fixed mean as random. The statement about exam scores describes individuals, whose spread is 3 points, far wider than this interval. The statement about sample means pins future means to an interval centred at one sample mean of 68, when they are centred at the unknown population mean.

The individual-scores reading is easy to dismiss with a number: exam scores have a standard deviation of 3, so about 90 percent of individual scores spread over nearly 10 points, not over a 1.64-point interval.

The probability reading is the subtle one. It sounds like the correct sentence, and in everyday speech people use it. In this course it is marked wrong, because once the interval is computed nothing in that statement is random any more.

38. Trap: σ in place of σ/√n

Trap

The trap

Example 8.2 again, 90 percent, sigma 3, n of 36, mean 68. A first attempt puts the population standard deviation straight into the error bound:

\[ \text{EBM} = 1.645 \times 3 = 4.935 \]

\[ (63.07,\ 72.94) \]

Wrong. That is about the width of the middle 90 percent of single exam scores.

The fix

The interval estimates a mean, so it uses the spread of sample means.

\[ \text{EBM} = 1.645 \times \frac{3}{\sqrt{36}} \]

\[ = 1.645 \times 0.5 = 0.8225 \]

\[ (67.18,\ 68.82) \]

The wrong interval is six times too wide, the square root of 36. A quick test catches it: an interval for a mean should be much narrower than the spread of the data. If your interval would cover most individual values, you used sigma where the standard error belongs.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 408 — the standard deviation must be the one appropriate to the parameter

39. What changes the width

Section

Part 4

40. Example 8.4: the exam interval at 95 percent

Worked example

Change Example 8.2 by using a 95 percent confidence level. Keep sigma 3, n of 36 and the sample mean 68, and find the new interval.

\[ \bar{x} = 68, \quad \sigma = 3, \quad n = 36, \quad \text{CL} = 0.95 \]

Find alpha and the tail area

Why: one minus the level, then halve it.

\[ \alpha = 1 - 0.95 = 0.05, \quad \frac{\alpha}{2} = 0.025 \]

The z-score

Why: 0.025 to the right means 0.975 to the left: invNorm(0.975, 0, 1).

\[ z_{0.025} = 1.96 \]

The error bound

Why: the standard error 0.5 has not changed.

\[ \text{EBM} = (1.96)\left(\frac{3}{\sqrt{36}}\right) = (1.96)(0.5) = 0.98 \]

Only one ingredient changed: the multiplier went from 1.645 to 1.96. So the error bound grows in exactly that ratio, from 0.8225 to 0.98. The sample, and so the centre, is the same.

Lower endpoint

Why: subtract.

\[ 68 - 0.98 = 67.02 \]

Upper endpoint

Why: add.

\[ 68 + 0.98 = 68.98 \]

Compare with Example 8.2: the centre is still 68, and each endpoint has moved out by 0.98 minus 0.8225, about 0.16 of a point. Same data, wider net.

In context: we estimate with 95 percent confidence that the true population mean of all statistics exam scores is between 67.02 and 68.98. Ninety-five percent of all intervals built this way contain the true mean.

Check the ratio of the error bounds

Why: the 95 percent bound over the 90 percent bound must equal the ratio of the z-scores.

\[ \frac{0.98}{0.8225} = 1.19 = \frac{1.96}{1.645} \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 413-414 — Example 8.4

41. More confidence, wider interval

Picture it

Put the exam intervals at five levels on one line. All come from the same 36 exams with mean 68; only the multiplier changes.

Figure (svg): Five interval bars stacked above a number line of mean exam score from 66.5 to 69.5, all centred at 68. From top to bottom they are the 80, 90, 95, 98 and 99 percent intervals, with half-widths 0.64, 0.82, 0.98, 1.16 and 1.29.

Same sample, same centre; only the level changes. Each step up in confidence buys it with width, and the 99 percent bar is twice as long as the 80 percent one.

The book's comparison: the 90 percent interval is (67.18, 68.82) and the 95 percent interval is (67.02, 68.98), which is wider. To be more confident that the interval contains the true mean, it has to be wider. A net cast wider is more likely to catch the fish.

\[ 0.98 > 0.8225 \quad \Rightarrow \quad (67.02,\ 68.98) \supset (67.18,\ 68.82) \]

The summary the book gives: increasing the confidence level increases the error bound, making the interval wider; decreasing the confidence level decreases the error bound, making it narrower. Each interval contains every interval of lower confidence from the same sample, because they share a centre.

Nothing about the data became more or less certain between the bars. You chose how often you are willing to be wrong, and the width followed from that choice.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 414 — Figure 8.5 and the summary: effect of changing the confidence level

42. Try It 8.4: pizza at 95 percent with 20 restaurants

Worked example

Return to the pizza delivery times: sigma 6 minutes, sample mean 36 minutes. Use a sample of 20 restaurants and find a 95 percent confidence interval for the true mean delivery time.

\[ \sigma = 6, \quad n = 20, \quad \bar{x} = 36, \quad \text{CL} = 0.95 \]

The z-score

Why: the 95 percent value.

\[ z_{0.025} = 1.96 \]

The standard error

Why: sigma over root 20.

\[ \frac{6}{\sqrt{20}} = \frac{6}{4.4721} = 1.3416 \]

The error bound

Why: z standard errors.

\[ \text{EBM} = (1.96)(1.3416) = 2.6296 \]

Two things widened this interval compared with Try It 8.2: the level rose from 90 to 95 percent, and the sample shrank from 28 to 20 restaurants. Both push the error bound up, so it has grown from 1.87 minutes to 2.63.

Lower endpoint

Why: subtract.

\[ 36 - 2.6296 = 33.37 \]

Upper endpoint

Why: add.

\[ 36 + 2.6296 = 38.63 \]

An interval more than five minutes wide is not very useful for a restaurant chain promising a delivery time. A narrower one would need either a larger sample or a lower confidence level, which is exactly the trade-off the next two examples measure.

In context: we estimate with 95 percent confidence that the true mean pizza delivery time is between 33.37 and 38.63 minutes.

Check the width against Try It 8.2

Why: the ratio of error bounds should be the ratio of z-scores times the root of the ratio of sample sizes.

\[ \frac{1.96}{1.645}\sqrt{\frac{28}{20}} = 1.191 \times 1.183 = 1.41 = \frac{2.6296}{1.8653} \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 414 — Try It 8.4

43. Example 8.5: changing the sample size

Worked example

Leave everything in Example 8.2 the same except the sample size, at the original 90 percent level. What happens to the error bound with n of 100 instead of 36, and with n of 25?

\[ \bar{x} = 68, \quad \sigma = 3, \quad z_{0.05} = 1.645 \]

n = 100: standard error

Why: root 100 is 10.

\[ \frac{3}{\sqrt{100}} = \frac{3}{10} = 0.3 \]

n = 100: error bound

Why: z standard errors.

\[ \text{EBM} = (1.645)(0.3) = 0.4935 \]

A bigger sample has a steadier mean, so the error bound falls from 0.8225 to 0.4935 and the interval narrows to (67.51, 68.49). Now go the other way.

n = 25: standard error

Why: root 25 is 5.

\[ \frac{3}{\sqrt{25}} = \frac{3}{5} = 0.6 \]

n = 25: error bound

Why: z standard errors.

\[ \text{EBM} = (1.645)(0.6) = 0.987 \]

A smaller sample has a jumpier mean, so the error bound rises to 0.987 and the interval widens to (67.01, 68.99). The book's summary: increasing the sample size decreases the error bound, making the interval narrower; decreasing it does the opposite.

Figure (svg): A curve of the error bound 1.645 times 3 over the square root of n, for n from 4 to 150, falling steeply at first and then flattening. Three points are marked: n of 25 with error bound 0.987, n of 36 with 0.822 and n of 100 with 0.494.

The curve falls like one over the square root of n. Going from 25 to 100, four times the sample, halves the error bound; going from 100 to 150 barely moves it.

Check the square-root law

Why: n of 100 is four times n of 25, so its error bound must be half.

\[ \frac{0.4935}{0.987} = 0.5 = \frac{1}{\sqrt{4}} \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 414-415 — Example 8.5 and the summary: effect of changing the sample size

44. Watch the interval shrink as n grows

Tweak it

The slider draws the sampling distribution of the mean exam score around the sample mean 68: a normal curve with standard error 3 over root n. The 90 percent interval is the stretch 1.645 standard errors either side of 68. Move n and watch where the curve's area goes.

Parameter explorer

Start at n equal to 9 and read off roughly where the curve meets the axis on each side. Then try 36 and 144. Each time n is multiplied by four, what happens to the width of the curve, and so to the 90 percent error bound?

\[ \text{EBM} = 1.645 \cdot \frac{3}{\sqrt{{n}}} \]

  • n — from 4 to 150: sample size n

The curve never moves off 68, because the centre is the sample mean, which this slider holds fixed. Only the spread responds to n, and only through its square root.

\[ n \to 4n \quad \Rightarrow \quad \text{EBM} \to \frac{\text{EBM}}{2} \]

That square root is the price list for precision. Halving the error bound costs four times the data, and cutting it to a third costs nine times. A researcher who wants a tight interval needs a lot more people, which is what Part 5's sample-size formula makes exact.

45. Try It 8.5: pizza with 50 restaurants

Worked example

Pizza delivery again: sample mean 36 minutes, sigma 6 minutes. The sample size is changed to 50 restaurants with the same sample mean. Find a 90 percent confidence interval for the population mean delivery time.

The z-score

Why: 90 percent.

\[ z_{0.05} = 1.645 \]

The standard error

Why: sigma over root 50.

\[ \frac{6}{\sqrt{50}} = \frac{6}{7.0711} = 0.8485 \]

The error bound

Why: z standard errors.

\[ \text{EBM} = (1.645)(0.8485) = 1.3958 \]

Compare with Try It 8.2, the same level and the same mean with 28 restaurants: the error bound there was 1.8653 minutes. Raising n to 50 has narrowed it by about a quarter, not by nearly half, because what matters is the square root of n.

Lower endpoint

Why: subtract.

\[ 36 - 1.3958 = 34.60 \]

Upper endpoint

Why: add.

\[ 36 + 1.3958 = 37.40 \]

In context: we estimate with 90 percent confidence that the true mean delivery time is between 34.60 and 37.40 minutes.

Check against Try It 8.2

Why: the error bounds should be in the ratio of the root of 28 to the root of 50.

\[ \sqrt{\frac{28}{50}} = 0.748 = \frac{1.3958}{1.8653} \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 415 — Try It 8.5

46. Rank five intervals by width

Ranking

Five intervals for the mean exam score, all with sigma 3, differ in level and sample size. Rank them from narrowest to widest. Think in terms of the two factors, z and root n, before computing anything.

Put in order

Order the intervals from narrowest to widest.

  1. 90% confidence, n = 100
  2. 95% confidence, n = 100
  3. 90% confidence, n = 36
  4. 95% confidence, n = 25
  5. 99% confidence, n = 36

Why: The error bounds are 1.645 times 0.3, 1.96 times 0.3, 1.645 times 0.5, 1.96 times 0.6 and 2.576 times 0.5, which are 0.49, 0.59, 0.82, 1.18 and 1.29. The two intervals from 100 exams are narrowest whatever their level, and the 99 percent interval from 36 exams is widest.

\[ 0.49 < 0.59 < 0.82 < 1.18 < 1.29 \]

A shortcut for comparisons like these: write each error bound as z times sigma over root n and compare the products, not the pieces. Neither factor alone decides the order; the 99 percent interval has the biggest z but not the smallest sample, and still ends up widest.

The comparison between the 95 percent interval from 25 exams and the 99 percent interval from 36 is the close one, and it needs the numbers: the smaller sample's standard error is 0.6 against 0.5, but the 99 percent multiplier is 2.576 against 1.96, and that wins.

Notice that sample size outweighed confidence level in most pairs here. A big sample at 95 percent beats a small sample at 90 percent, so a researcher who wants both high confidence and a narrow interval has to pay for it in data.

47. What moves the error bound, and what does not

Comparison

The error bound has three ingredients: the z from the level, sigma, and n. Fill in what each change does to the error bound, and then to where the interval sits. Use the words larger, smaller or same.

Comparison matrix

change (all else fixed)zσ/√nEBMcentre of interval
level 90% → 95%largersamelargersame
n 36 → 100samesmallersmallersame
σ 3 → 6samelargerlargersame
x̄ 68 → 70samesamesamemoves up 2

The last row is the one people get wrong. A different sample mean slides the whole interval along the line but does not change its width at all, because the sample mean is not in the error bound. When sigma is known, the width is decided before you collect any data.

\[ \text{EBM} = z_{\alpha/2}\,\frac{\sigma}{\sqrt{n}} \quad \text{(no } \bar{x} \text{ anywhere)} \]

The third row deserves a sentence too. Doubling sigma doubles the standard error, so it doubles the error bound: a population twice as spread out needs an interval twice as wide for the same confidence and sample size. You cannot change sigma, but it tells you which studies will be expensive.

That fact is what makes Part 5 possible: since the width does not depend on the data, you can choose n in advance to get the width you want.

48. Why not 100 percent confidence?

Counterexample

If more confidence is better, why does anyone settle for 90 or 95 percent? Push the level toward 100 percent with the exam scores and see what the interval turns into.

Discussion prompt

For the exam scores (sample mean 68, standard error 0.5), the 99.9 percent z is 3.291 and the 99.99 percent z is 3.891. Find both intervals. What would a 100 percent interval have to be, and why is it useless?

So confidence and precision pull against each other. The certain interval is useless, and a very narrow interval is often wrong. The conventional 90, 95 and 99 percent levels are a compromise most fields have settled on, and the person building the interval chooses among them.

The real way to get an interval that is both narrow and highly confident is not to raise the level but to raise n, and that costs money and time. That is the trade-off Part 5 turns into a formula.

49. Working backwards, and sample size

Section

Part 5

50. Taking a published interval apart

Concept

A study often reports only the interval. Because every interval in this section is the sample mean plus and minus the same error bound, the interval carries both numbers inside it, and you can recover them.

Figure (svg): A number line from 41.5 to 48.5 with an interval bar from 42.12 to 47.88. A yellow dot at its midpoint is labelled x-bar equals midpoint equals 45. Orange arrows above run from the midpoint to each end, each labelled EBM equals 2.88.

A symmetric interval carries its own centre and its own half-width. The midpoint is the sample mean; half the length is the error bound.

The book gives two methods for each, so you can use whichever matches what you know.

\[ \text{EBM} = \text{upper} - \bar{x} \quad \text{or} \quad \text{EBM} = \frac{\text{upper} - \text{lower}}{2} \]

\[ \bar{x} = \text{upper} - \text{EBM} \quad \text{or} \quad \bar{x} = \frac{\text{upper} + \text{lower}}{2} \]

Why these work: the upper endpoint is the mean plus the error bound, so subtracting the mean leaves the bound. The two endpoints are the mean plus and minus the bound, so their difference is twice the bound and their sum is twice the mean.

Once you have the error bound, you can go further. If you also know sigma and the level, the error bound formula has only n left in it, so you can recover the sample size the study used. The own example after Example 8.6 does exactly that.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 415 — finding the error bound and the sample mean

51. Example 8.6: recovering the error bound and the mean

Worked example

A confidence interval is (67.18, 68.82). Find the error bound, first knowing the sample mean is 68, then without knowing it. Then find the sample mean both ways.

EBM, mean known

Why: upper endpoint minus the sample mean.

\[ \text{EBM} = 68.82 - 68 = 0.82 \]

EBM, mean unknown

Why: half the length of the interval.

\[ \text{EBM} = \frac{68.82 - 67.18}{2} = \frac{1.64}{2} = 0.82 \]

Two routes, one answer, which is the point of the symmetric form. The second route is the one you need when a report gives only the interval.

Mean, EBM known

Why: upper endpoint minus the error bound.

\[ \bar{x} = 68.82 - 0.82 = 68 \]

Mean, EBM unknown

Why: the midpoint of the endpoints.

\[ \bar{x} = \frac{68.82 + 67.18}{2} = \frac{136}{2} = 68 \]

Which route should you use? Whichever needs only what you were given. A news report usually gives the two endpoints and nothing else, so the half-length and midpoint routes are the ones you will use most; the subtraction routes are quicker when one of the two numbers is already known.

These are the numbers of Example 8.2, rounded, so the method has given back what that example put in. Notice that the rounding shows: the true error bound was 0.8225, and from the rounded endpoints you can only recover 0.82.

Check by rebuilding the interval

Why: the recovered mean and bound must reproduce both endpoints.

\[ 68 - 0.82 = 67.18, \quad 68 + 0.82 = 68.82 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 415-416 — Example 8.6

52. Try It 8.6: take (42.12, 47.88) apart

Fill the middle

A confidence interval is (42.12, 47.88). Find the error bound and the sample mean, using the methods that need only the two endpoints.

Fill in the blanks

\text2.88 = \frac45___ = ___, \qquad \bar___ = \frac______ = ___

Why: The interval is 5.76 long, so half of it, 2.88, is the error bound. The endpoints add to 90, so the midpoint, 45, is the sample mean.

Then check by rebuilding: 45 minus 2.88 is 42.12 and 45 plus 2.88 is 47.88, the two endpoints you started with.

\[ 45 - 2.88 = 42.12, \qquad 45 + 2.88 = 47.88 \]

Working backwards is also a fast way to catch an error in someone else's interval. If a report says the sample mean was 45 but the interval is (42.0, 47.0), the midpoint is 44.5, so either the mean or the interval is wrong, or the interval is not the symmetric kind this section builds.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 416 — Try It 8.6

53. Recovering the sample size from an interval

Worked example

An own example, in the spirit of the homework. A report gives the 95 percent interval (67.02, 68.98) for mean exam score and says the population standard deviation is 3 points, but not how many exams were sampled. Find n.

\[ (67.02,\ 68.98), \quad \text{CL} = 0.95, \quad \sigma = 3 \]

Recover the error bound

Why: half the length.

\[ \text{EBM} = \frac{68.98 - 67.02}{2} = \frac{1.96}{2} = 0.98 \]

Write the error bound formula

Why: with the 95 percent z.

\[ 0.98 = 1.96\,\frac{3}{\sqrt{n}} \]

Now n is the only unknown. Solve for the square root of n first, then square: two separate moves, because squaring too early tangles the algebra.

Solve for root n

Why: multiply both sides by root n and divide by 0.98.

\[ \sqrt{n} = \frac{1.96 \times 3}{0.98} = \frac{5.88}{0.98} = 6 \]

Square

Why: root n is 6.

\[ n = 6^{2} = 36 \]

This is Example 8.4's interval, built from 36 exams, and the method has recovered the 36. With rounded endpoints in a real report the answer may come out a little off a whole number; round to the nearest whole number, since the study used a whole number of subjects.

Check by rebuilding the error bound

Why: 36 exams must give back 0.98.

\[ 1.96 \times \frac{3}{\sqrt{36}} = 1.96 \times 0.5 = 0.98 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 414 — the interval of Example 8.4

54. "Within E of the mean" is an error bound

Concept

Sample-size problems never say error bound. Example 8.7 says it wants to be 95 percent confident that the sample mean age is within two years of the true mean age. Translate that phrase before you reach for a formula.

\[ |\bar{x} - \mu| \le 2 \quad \Longleftrightarrow \quad \mu \in (\bar{x} - 2,\ \bar{x} + 2) \]

The sample mean within two years of mu is the same event as mu within two years of the sample mean, which is the same as the interval with error bound 2 containing mu. It is the flip from Part 1 again, used the other way round: there you had a sample and wanted the interval; here you have the interval width you want and need the sample.

the problem saysit means
within two years of the true meanEBM = 2 years
a margin of error of 3 pointsEBM = 3 points
a 95% margin of error of 1 inchEBM = 1 inch, CL = 0.95
an interval no wider than 4 unitsEBM = 2: the whole width is twice the bound

The last row is a trap of its own. If a problem fixes the total width of the interval, the error bound is half of it; plugging the full width into the formula gives a sample four times too small.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 416 — Example 8.7: within two years of the true population mean

55. Solving the error bound for n

Intuition

A researcher planning a study chooses the confidence level and the error bound they can live with, and needs to know how many subjects that takes. The error bound formula already links those three things to n; solve it for n.

\[ \text{EBM} = z\,\frac{\sigma}{\sqrt{n}} \]

Multiply both sides by the square root of n, so it is no longer in a denominator.

\[ \text{EBM}\,\sqrt{n} = z\,\sigma \]

Divide both sides by the error bound to isolate the square root of n.

\[ \sqrt{n} = \frac{z\,\sigma}{\text{EBM}} \]

Square both sides to get n itself.

\[ n = \frac{z^{2}\sigma^{2}}{\text{EBM}^{2}} \]

Read off what drives the sample size. It grows with the square of z, so higher confidence is expensive. It grows with the square of sigma, so a more variable population needs more data. And it grows with one over the square of the error bound, so halving the margin multiplies the sample by four.

Here z is z sub alpha over two for the chosen level, and the formula needs sigma known in advance. In practice that often comes from an earlier study or a pilot sample.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 416 — calculating the sample size n

56. Example 8.7: how many Foothill College students?

Worked example

The population standard deviation for the age of Foothill College students is 15 years. To be 95 percent confident that the sample mean age is within two years of the true population mean age, how many randomly selected students must be surveyed?

\[ \sigma = 15, \quad \text{EBM} = 2, \quad \text{CL} = 0.95 \]

The z-score

Why: 95 percent.

\[ z = z_{0.025} = 1.96 \]

Write the sample size formula

Why: with the numbers in place.

\[ n = \frac{z^{2}\sigma^{2}}{\text{EBM}^{2}} = \frac{(1.96)^{2}(15)^{2}}{(2)^{2}} \]

Square each piece

Why: 1.96 squared, 15 squared, 2 squared.

\[ n = \frac{(3.8416)(225)}{4} \]

Multiply

Why: the numerator.

\[ n = \frac{864.36}{4} \]

Divide

Why: the book's value.

\[ n = 216.09 \]

You cannot survey 0.09 of a student, so the answer must be a whole number, and it must be rounded up, to 217. Rounding down to 216 leaves the error bound just above 2 years, so the study would miss its own target.

Figure (svg): A curve of the required sample size, 1.96 times 15 over E all squared, for target error bounds E from 1 to 5 years. It falls from 864 at E of 1 to 35 at E of 5. A point is marked at E of 2 with n of 216.09.

Halving the target error bound from 2 years to 1 multiplies the sample by four, from 217 to 865. Precision is paid for in the square.

In context: 217 Foothill College students should be surveyed to be 95 percent confident that the sample mean age is within two years of the true population mean age.

Check by putting n = 217 back in

Why: the error bound must be at most 2 years.

\[ 1.96 \times \frac{15}{\sqrt{217}} = 1.96 \times 1.0183 = 1.996 \le 2 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 416 — Example 8.7

57. Trap: rounding the sample size to the nearest whole number

Trap

The trap

Example 8.7 gives n equal to 216.09. Ordinary rounding says:

\[ n = 216.09 \approx 216 \]

Wrong. With 216 students the error bound is:

\[ 1.96 \times \frac{15}{\sqrt{216}} = 2.0004 \]

That is just over the target of 2 years.

The fix

The formula gives the smallest n that exactly meets the target. Any whole number below it falls short, so always round up.

Figure (svg): Seven points of the error bound 1.96 times 15 over root n for n from 214 to 220, falling from 2.0097 to 1.9821, with a dashed red line at the target 2. The point at n of 216, 2.0004, is just above the line; the point at 217, 1.9958, is just below it.

Every n at or below 216 gives an error bound above 2 years. The first whole number that meets the target is 217, which is why the answer is always rounded up.

\[ n = 217 \Rightarrow \text{EBM} = 1.9958 \le 2 \]

The rule holds even when the decimal is tiny: 216.09 becomes 217, and 34.57 becomes 35. The book says it in capitals: always round the answer UP to the next higher integer to ensure the sample size is large enough.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 416 — Example 8.7: always round up

58. Try It 8.7: basketball players' heights

Worked example

The population standard deviation for the height of high school basketball players is 3 inches. To be 95 percent confident that the sample mean height is within one inch of the true population mean height, how many randomly selected players must be measured?

\[ \sigma = 3, \quad \text{EBM} = 1, \quad z = 1.96 \]

Substitute

Why: into the sample size formula.

\[ n = \frac{(1.96)^{2}(3)^{2}}{(1)^{2}} \]

Square each piece

Why: 1.96 squared is 3.8416, 3 squared is 9.

\[ n = \frac{(3.8416)(9)}{1} \]

Multiply

Why: the denominator is 1.

\[ n = 34.5744 \]

Round up

Why: to the next whole number.

\[ n = 35 \]

The answer 34.57 is closer to 35 than to 34, so ordinary rounding and rounding up agree here. They will not always agree, as Example 8.7 showed, so make rounding up a fixed habit rather than a judgment call each time.

In context: 35 randomly selected high school basketball players should be measured to be 95 percent confident that the sample mean height is within one inch of the population mean height. It is a much smaller study than Foothill's, because sigma is small compared with the error bound wanted: the ratio is 3 to 1 here and 15 to 2 there.

Check both neighbours

Why: 35 must meet the target and 34 must not.

\[ \frac{1.96 \times 3}{\sqrt{35}} = 0.994 \le 1, \quad \frac{1.96 \times 3}{\sqrt{34}} = 1.008 > 1 \quad \checkmark \]

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 416 — Try It 8.7

59. Pushing the Foothill study to its limits

Edge cases

Example 8.7 needed 217 students for a two-year margin at 95 percent. A dean asks for more: first a tighter margin, then more confidence. Before computing, guess how big each change makes the study.

Discussion prompt

Using sigma of 15 years, find the sample size for (a) a one-year margin at 95 percent, (b) a half-year margin at 95 percent, and (c) a two-year margin at 99 percent. Which request is the most expensive, and why?

The margin enters the formula squared in the denominator, so every halving of the margin quadruples the sample; the confidence level enters only through z, which grows slowly from 1.96 to 2.576. That asymmetry is why polls quote margins of about three percentage points rather than one: the last two points would cost nine times the interviews.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 416 — the sample size formula

60. The collaborative exercise: meals eaten out

Real world

The chapter opens with a class activity: record the number of meals each student ate out last week, assume the standard deviation is three meals, and build an approximate 95 percent interval using two standard errors. Here is an invented class of 20, so you can do it end to end.

meals eaten out last week, 20 students
3527461384
2563492543

Discussion prompt

Using these 20 values and sigma of 3 meals, find the sample mean, the error bound with two standard errors, and the approximate 95 percent interval. Then write the book's sentence. Would the interval apply to all students at the college?

The exercise shows the two conditions this section quietly relies on: a random sample, and a known sigma. Here sigma is assumed, which is rarely realistic, and that is the problem Section 8.2 solves with a new curve, Student's t distribution.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8 Introduction, pp. 406-407 — Collaborative Exercise

61. Putting it together

Section

Part 6

62. Pattern: six steps to an interval

Pattern

Every interval in this section came from the same five calculations and one sentence. The book lists them as the steps to construct and interpret a confidence interval.

Figure (svg): Six boxes in a row joined by arrows: the sample mean from the data; the confidence level converted to alpha; the z-score with area 1 minus alpha over two to its left; the error bound z times sigma over root n; the interval x-bar plus or minus the error bound; and the sentence naming the level, mu and units, highlighted in yellow.

Five steps of arithmetic and one of language. The sixth step is where most marks are lost.
  1. Calculate x̄ from the sample data; σ is already known in this section.
  2. Convert the level: α = 1 − CL, with α/2 in each tail.
  3. Find z with area 1 − α/2 to its left: 1.645, 1.96, 2.326 or 2.576 for the usual levels.
  4. Calculate EBM = z · σ/√n, with the standard error, never σ alone.
  5. Construct the interval (x̄ − EBM, x̄ + EBM).
  6. Interpret: level, the population mean in context, both endpoints, units.

\[ \bar{x} \pm z_{\alpha/2}\,\frac{\sigma}{\sqrt{n}} \]

Two reversals use the same formula. Given an interval, its midpoint is the sample mean and half its length is the error bound. Given a target error bound, solving for n gives the sample size, always rounded up.

And two checks catch most errors: a higher level must have a larger z, and the midpoint of your interval must be your sample mean.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, p. 408 — the steps to construct and interpret the interval

63. Check: building a 98 percent interval

Check

A fresh set of numbers. Decide the z, the standard error and the error bound in that order before choosing.

Check your understanding

A population has standard deviation 4. A random sample of 64 has mean 50. Which is the 98 percent confidence interval for the population mean?

  • A. (48.84, 51.16) (correct)
  • B. (48.97, 51.03)
  • C. (40.70, 59.30)
  • D. (49.85, 50.15)

Answer: A

Why: Alpha is 0.02, so each tail holds 0.01 and the z is invNorm of 0.99, which is 2.326. The standard error is 4 over the square root of 64, which is 0.5. The error bound is 2.326 times 0.5, about 1.163, so the interval runs from 48.84 to 51.16.

Why B tempts people
That uses invNorm of 0.98, which is 2.054: alpha was not split between the two tails, so the interval is too narrow.
Why C tempts people
That multiplies the z by sigma itself, 4, instead of by the standard error, giving an interval eight times too wide.
Why D tempts people
That divides sigma by n, 64, instead of by its square root, which makes the standard error far too small.

\[ z_{0.01} = 2.326, \quad \frac{4}{\sqrt{64}} = 0.5, \quad \text{EBM} = 1.163, \quad (48.84,\ 51.16) \]

Check the answer the usual two ways. The midpoint of 48.84 and 51.16 is 50, the sample mean. And the 98 percent z, 2.326, sits between the 95 percent value 1.96 and the 99 percent value 2.576, as a 98 percent level must.

In context, if this were a measurement: we estimate with 98 percent confidence that the population mean is between 48.84 and 51.16. Each wrong option is one of this lesson's traps: alpha not halved, sigma in place of the standard error, and n in place of its square root.

64. Check: what 95 percent confident means

Check

This check is about the sentence, not the number. Read each option for what it claims is random and what it describes.

Check your understanding

A 95 percent confidence interval for the mean exam score is (67.02, 68.98). Which statement correctly explains the 95 percent?

  • A. About 95% of intervals built this way from repeated samples contain the true mean exam score. (correct)
  • B. There is a 95% probability that the true mean exam score is between 67.02 and 68.98.
  • C. About 95% of all exam scores are between 67.02 and 68.98.
  • D. About 95% of samples of 36 exams will have a mean between 67.02 and 68.98.

Answer: A

Why: The confidence level describes the method: across many samples, about 95 percent of the intervals produced this way contain the fixed population mean. This particular interval either contains it or does not.

Why B tempts people
That treats the population mean as random. It is fixed, and so are these endpoints, so the event has no 95 percent probability; the book calls the level the percent of intervals that contain the parameter.
Why C tempts people
That describes individual scores, which vary with standard deviation 3 points; the middle 95 percent of them spans nearly 12 points, not 2.
Why D tempts people
Future sample means are centred at the unknown population mean, not at 68, so this interval does not describe where they fall.

Each wrong statement swaps the subject: the fixed parameter treated as random, individual exams in place of the mean, or future sample means in place of the population mean. The correct one is about the procedure.

A safe pair of sentences: we estimate with 95 percent confidence that the true mean exam score is between 67.02 and 68.98; ninety-five percent of intervals built this way contain the true mean.

65. Check: planning a study

Check

Use the sample size formula, then decide how to round.

Check your understanding

A population has standard deviation 12. How large a sample is needed to estimate the mean to within 3 units with 90 percent confidence?

  • A. 44 (correct)
  • B. 43
  • C. 62
  • D. 7

Answer: A

Why: With z equal to 1.645, n is 1.645 times 12 over 3, all squared: 6.58 squared, which is 43.29. Rounding up gives 44, the smallest whole number whose error bound is at most 3.

Why B tempts people
That rounds 43.29 down. With 43 the error bound is just above 3, so the target is missed.
Why C tempts people
That uses the 95 percent z, 1.96, instead of 1.645 for 90 percent.
Why D tempts people
That forgets to square: 1.645 times 12 over 3 is the square root of n, not n.

\[ n = \left(\frac{1.645 \times 12}{3}\right)^{2} = (6.58)^{2} = 43.29 \Rightarrow 44 \]

Verify by putting 44 back in: 1.645 times 12 over the square root of 44 is 2.98, just inside the target of 3. With 43 it is 3.01, just outside, which is the whole case for rounding up.

In context: the study needs 44 subjects to be 90 percent confident of estimating the mean within 3 units. The two errors to guard against are the ones the formula's shape invites, forgetting the square and rounding the wrong way.

66. Explain 95 percent confident to someone else

Explain it to yourself

Being able to say what a confidence level means, without saying anything false, is the real test of this lesson. Try it in plain words before reading the model answer.

Discussion prompt

A friend reads that a 95 percent confidence interval for mean exam score is (67.02, 68.98) and says: so there is a 95 percent chance the mean is in there. In a few sentences, explain what is wrong with that, what the 95 percent does mean, and why the 95 percent interval is wider than the 90 percent interval (67.18, 68.82).

A strong explanation names what is fixed (the mean), what is random (the interval, before sampling), and what the percentage counts (intervals over repeated samples). Then the width comparison follows from the z-scores, not from the data.

If your explanation used the word chance about this specific interval, rewrite it so the chance belongs to the method. That one change is what most test questions on this topic are checking.

67. Exit ticket: a reported commute interval

Exit ticket

A city report gives a 95 percent confidence interval of (24.55, 29.45) minutes for the mean commute time of its workers. The population standard deviation is 10 minutes.

Discussion prompt

Find (a) the sample mean and the error bound, (b) the sample size the city used, and (c) the 90 percent interval from the same sample. Write one sentence interpreting (c).

The marks are in three places: taking the interval apart by midpoint and half-length, solving the error bound for n without squaring too early, and keeping the sample mean in the centre when the level changes.

A quick self-check: the 90 percent interval must sit inside the 95 percent one, and 24.94 to 29.06 does sit inside 24.55 to 29.45.

68. Recap

Recap

A point estimate is one number, the sample mean; a confidence interval surrounds it with an error bound on each side. The interval is the random thing and the population mean is fixed, so the confidence level describes the method across repeated samples.

pieceformulaExample 8.2
α and the tailsα = 1 − CL, α/2 each side0.10, 0.05
zinvNorm(1 − α/2)1.645
standard errorσ/√n3/6 = 0.5
error boundEBM = z · σ/√n0.8225
interval(x̄ − EBM, x̄ + EBM)(67.18, 68.82)
sample sizen = z²σ²/EBM², round upFoothill: 217

Raising the confidence level raises z and widens the interval; raising n shrinks the standard error by its square root and narrows it. A published interval gives back its sample mean as the midpoint and its error bound as half the length.

Every interval ends in a sentence: we estimate with so much confidence that the true population mean, in context, is between two numbers, in units. Introductory Business Statistics builds the same interval with business data, if you want more practice.

OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution §8.1, pp. 406-416 — the whole section

OpenStax Introductory Business Statistics 2e, §8.1 A Confidence Interval When the Population Standard Deviation Is Known or Large Sample Size §8.1 A Confidence Interval When the Population Standard Deviation Is Known or Large Sample Size — the same interval in business contexts

Sources

  1. OpenStax Introductory Statistics 2e, §8.1 A Single Population Mean using the Normal Distribution — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0, pp. 405-417
  2. OpenStax Introductory Business Statistics 2e, §8.1 A Confidence Interval When the Population Standard Deviation Is Known or Large Sample Size — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0

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

Book on Wyzant · Text (657) 465-8108