8.2 A Single Population Mean using the Student t Distribution

Section 8.1 assumed the population standard deviation was known, which in practice it almost never is. Replacing sigma with the sample standard deviation adds a second source of uncertainty, and for small samples the normal distribution no longer describes the resulting statistic. William Gosset, working at the Guinness brewery where experiments yielded very few samples, found the distribution that does, and published it under the pen name Student. The t distribution is symmetric about zero like the normal but has more probability in its tails and less in its centre, and its exact shape depends on the degrees of freedom, which equal the sample size minus one because the deviations used to compute the sample standard deviation must sum to zero. As the degrees of freedom grow the t curve approaches the normal, so the price of not knowing sigma is large for a small sample and negligible for a large one.

Subject: Statistics · 64 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 Student t Distribution

Title

Statistics · Chapter 8 — Confidence Intervals

When sigma is unknown, estimate it, and pay for the estimate with a wider interval

2. What this lesson gives you

Objectives

Section 8.1 built a confidence interval for a population mean on one awkward assumption: that you already know the population standard deviation sigma. You almost never do. If you don't know the mean, you rarely know how spread out the population is either. This section removes that assumption.

The arithmetic barely changes: the interval is still the sample mean plus or minus an error bound. Two things in the error bound change. The spread is now the sample's own s, and the multiplier comes from a new family of curves, Student t distributions. Most of this lesson is about why that second change is needed and how big it is.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 416-420 — the section this deck covers

3. Before anything new: a Section 8.1 interval

Warm-up

Recall how Section 8.1 built an interval. A sample of 36 has mean 68, and the population standard deviation is known to be 3. Build the 90 percent confidence interval.

Discussion prompt

Find the error bound and the 90 percent confidence interval for the population mean, with sample mean 68, sigma 3 and n equal to 36.

Look at what that calculation needed: a z-score, the sample size, the sample mean, and sigma. Now imagine the problem had said nothing about sigma, and you had only the 36 data values. You could compute their standard deviation s and put it where sigma was. The question for this whole section is whether that substitution is honest, and what it costs.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.1, pp. 415-416 — the error bound and Example 8.6

4. Why a different distribution

Section

Part 1

5. Sigma is almost never known

Concept

The book opens the section with a plain fact: in practice we rarely know the population standard deviation. Section 8.1's examples supplied it, from years of records or a manufacturer's specification, but a real study usually has only its own data.

For large samples this was never much of a problem. Statisticians used the sample standard deviation s as an estimate for sigma and carried on, and the intervals came out close enough. The trouble came with small samples, where the intervals built that way were inaccurate: they missed the true mean more often than their stated confidence level promised.

William Gosset ran into exactly this at the Guinness brewery in Dublin. His experiments with hops and barley produced very few samples, and simply replacing sigma with s did not give accurate intervals. He realised the normal distribution was the wrong curve for the calculation, and that the right one depended on the sample size.

Guinness did not let its staff publish under their own names, so in 1908 he wrote as Student, and the distribution he found is still called Student's t distribution. The name sounds like a classroom simplification; it is a pen name.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 416-417 — Gosset and Student's t-distribution

6. The t-score: s where sigma was

Concept

Section 8.1 measured how far a sample mean landed from the population mean in standard errors, using sigma:

\[ z = \frac{\bar{x} - \mu}{\sigma/\sqrt{n}} \]

When sigma is unknown, put the sample standard deviation in its place. The result is called a t-score:

\[ t = \frac{\bar{x} - \mu}{s/\sqrt{n}} \]

The book is careful to say the t-score has the same interpretation as the z-score: it measures how far the sample mean is from its mean, the population mean, in units of the estimated standard error. A t-score of 2 still means two standard errors above.

What changes is its distribution. If the population is approximately normal and the sample is simple random, these t-scores follow a Student's t distribution with n minus 1 degrees of freedom, and for each sample size there is a different t distribution.

Student's t distribution — The distribution of the t-score, the standardised sample mean with s in place of sigma, when sampling from an approximately normal population. There is one curve for each number of degrees of freedom, df equal to n minus 1.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — the t-score and its distribution

7. How often does a t-score pass 1.96?

Prediction

For a z-score, 1.96 is the famous cut-off: only 5 percent of z-scores land further than 1.96 from zero. Now take tiny samples of four from a normal population and compute the t-score for each, using each sample's own s.

Decide before you look at the simulation on the next slide. Think about what happens to the fraction when the denominator, s over root n, can come out too small as well as too large.

Predict first

With samples of size four, roughly what fraction of t-scores land further than 1.96 from zero?

  • About 5 percent, the same as for z
  • About 15 percent
  • About 2 percent, fewer than for z
  • About 50 percent

Correct: About 15 percent

Why: With three degrees of freedom the tails of the t distribution are much heavier than the normal's, and the exact chance of passing 1.96 either way is 0.145. Samples whose s happens to be small give inflated t-scores, and there are no samples that make up for it on the other side.

If you chose 5 percent, you assumed s is as good as sigma. If you chose fewer, you may have reasoned that s is sometimes too big and sometimes too small, so the errors cancel. They don't cancel, and the next two slides show why.

8. The simulation: tails that are too heavy for the normal

Intuition

Here are 3000 samples, each of four values drawn from a normal population. For each, the t-score uses the sample's own s. The grey and yellow bars are those t-scores; the dashed orange curve is the standard normal that Section 8.1 would use.

Figure (svg): A histogram of 3000 simulated t-scores, each from a sample of four values drawn from a normal population, in bins half a unit wide from minus 6 to 6. A dashed orange standard normal curve is drawn over it; the bars are lower than the curve near zero and higher than it beyond 2 on each side. A blue curve, the t distribution with 3 degrees of freedom, follows the bars. 452 of the t-scores lie beyond plus or minus 1.96, against 147 of the matching z-scores.

Divide by a known sigma and about 5 percent of scores pass 1.96. Divide by each sample's own s and nearly three times as many do: the tails have thickened.

Compare the bars with the dashed curve. Near zero the bars fall short of it: the t-scores are less concentrated in the middle. Beyond 2 on either side the yellow bars stand well above it. Of the 3000 t-scores, 452 landed beyond plus or minus 1.96, about 15 percent, where the matching z-scores managed only 147, about 5 percent.

The solid blue curve is Gosset's answer, the t distribution with 3 degrees of freedom, one fewer than the sample size. It follows the bars closely, including the heavy tails. That is what the book means by thicker in the tails and shorter in the centre than the standard normal.

\[ P(|T| > 1.96) = 0.145 \text{ for } T \sim t_{3}, \qquad P(|Z| > 1.96) = 0.05 \]

So an interval built with 1.96 and s from four observations would miss the true mean about 15 percent of the time while claiming 5. That is the inaccuracy Gosset found at the brewery.

9. Why the tails thicken: a small s inflates t

Picture it

Where do the extra large t-scores come from? Plot each sample twice over: across, its standard deviation s; up, how far its mean landed from the population mean. Here are 400 samples of four from a population with sigma equal to 1.

Figure (svg): A scatter of 400 samples of size four from a normal population with sigma 1. The horizontal axis is the sample standard deviation s, from 0 to 2.5; the vertical axis is the sample mean minus mu. Two dashed horizontal lines at plus and minus 0.98 mark where a z-score reaches 1.96. Two solid lines fanning out from the origin mark where a t-score reaches 1.96. 42 red points sit inside the horizontal band but outside the fan: samples whose mean was ordinary but whose s was small.

A z-score only cares how far the mean landed. A t-score also divides by s, so a sample with an unluckily small s gets a big t even when its mean is ordinary.

A z-score passes 1.96 when the point is above or below the dashed orange lines, because its denominator is fixed. A t-score passes 1.96 when the point is outside the solid blue fan, because its denominator grows with s. Near the left edge, where s is small, the fan is narrow.

The 42 red points are the story. Their means landed an ordinary distance from the truth, inside the orange band, but their s came out small, so dividing by s over root n made the t-score large. A large s does the reverse, shrinking t, but it cannot pull a t-score back from the tails; it only pushes moderate scores towards the middle.

So a t-score carries two sources of chance where a z-score carries one: where the sample mean landed, and how big s happened to be. The book's phrase for this is that the spread of the t distribution is greater than the spread of the standard normal.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — more probability in the tails

10. s is a random variable too

Concept

The wedge showed that small values of s make large t-scores. How often does s come out small? That depends on the sample size, because s is a statistic: it changes from sample to sample, just as the sample mean does, and it has its own distribution.

Figure (svg): Two overlaid histograms of the sample standard deviation s, each from 2000 simulated samples drawn from a normal population with sigma equal to 1, for s from 0 to 2.5. The orange histogram, samples of 4, is wide and spreads from near 0 to beyond 2. The blue histogram, samples of 15, is tall and narrow, crowded between about 0.6 and 1.4. A dashed yellow line marks sigma equal to 1. A shaded band below 0.5 holds 260 of the samples of 4 and 3 of the samples of 15.

The true sigma is 1. From four observations, s is often far from it, and below 0.5 surprisingly often; from fifteen, s rarely strays far.

Here the true sigma is exactly 1. From samples of four, s scattered from almost zero to well past 2, and 260 of the 2000 samples, about 13 percent, gave an s below one half. From samples of fifteen only 3 did. A larger sample pins s down close to sigma.

Put that together with the t-score's denominator. When s can easily come out at half of sigma, the estimated standard error can easily be half the true one, and the t-score twice as large as the z-score would have been. That is where the heavy tails come from, and it is why the damage is worst for small samples.

\[ \frac{s}{\sqrt{n}} \text{ too small} \;\Rightarrow\; t \text{ too large} \;\Rightarrow\; \text{heavy tails} \]

It also explains why the t distribution needs a separate curve for every sample size. The amount of wobble in s depends on n, so the amount of extra spread in t depends on n too. Gosset's insight was that this extra spread can be calculated exactly, if the population is normal.

11. Reading the t-score piece by piece

Notation

Every piece of the t-score is something you can compute from data except one, the population mean, which is the thing you are trying to pin down.

Annotate

On: \( t = \frac{\bar{x} - \mu}{s/\sqrt{n}}, \qquad T \sim t_{n-1} \)

  • How far the sample mean landed from the population mean. The numerator is the same as in a z-score.
  • The sample standard deviation, computed from the same data. It replaces σ and varies from sample to sample.
  • The estimated standard error: how much sample means of this size typically vary.
  • The t-scores follow a Student's t distribution whose degrees of freedom are one fewer than the sample size.

Hover over each piece. The pieces are the same as in a z-score except one, and that one piece is the whole reason for this section.

Try it on the acupuncture data you will meet in Part 3, where the sample mean is 8.2267, s is 1.6722 and n is 15. If someone claimed the population mean were exactly 8, the sample mean would sit about half a standard error above it:

\[ t = \frac{8.2267 - 8}{1.6722/\sqrt{15}} = \frac{0.2267}{0.4318} = 0.525 \]

That reads exactly as a z-score would: the sample mean is about half an estimated standard error above 8, an entirely ordinary distance. The difference is only which curve tells you how ordinary.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — t = (x̄ − μ)/(s/√n) and T ~ t_df

12. The rule today: t whenever s stands in for sigma

Concept

Older textbooks give a different rule, and you will meet it, so it is worth knowing where it came from. Until the mid-1970s some statisticians used the normal distribution for large samples and the t distribution only for samples of at most 30.

That rule was a convenience. Before calculators, every t value came from a printed table with a row for each df, and for large samples the t values were so close to the normal's that the table hardly mattered. Thirty was a rough point past which the difference looked small enough to ignore.

With graphing calculators and computers, that convenience is gone, and the book states the modern practice plainly: use Student's t distribution whenever s is used as an estimate for sigma.

\[ \sigma \text{ known} \;\Rightarrow\; z, \qquad s \text{ used for } \sigma \;\Rightarrow\; t \]

Notice what does not appear in that rule: the sample size. The deciding question is never how big the sample is. It is whether the standard deviation in the problem was handed to you as a population fact, or computed from the sample in front of you. For a large sample the two answers barely differ, so following the modern rule costs nothing and avoids a judgment call.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — the modern practice

13. z or t? Read where the spread came from

Sorting

For each situation, find the standard deviation in the problem and ask where it came from. A value described as known, or as the population standard deviation, means z. A value computed from the sample means t, whatever the sample size.

Sort into buckets

Sort each situation by the distribution its confidence interval should use.

z: σ is known
σ = 3 is known; a sample of 36 has mean 68; 20 heads of lettuce with s = 0.1 lb, and the population σ is known to be 0.2 lb; Foothill College ages: σ has been consistently 15; 25 students sampled
t: s estimates σ
15 acupuncture subjects; s = 1.6722 computed from their rates; 108 Americans' monthly TV hours; s = 32 hours from the survey; 70 ER patients; sample standard deviation 0.5 hours
z
In each of these the population standard deviation is stated as a known fact, so Section 8.1's normal method applies. For the lettuce, the known sigma is used and the sample's s is simply not needed.
t
In each of these the only spread available is computed from the sample, so the t distribution is correct, even for 70 or 108 people.

The lettuce is the one to slow down on: it gives both numbers. When sigma is known, use it; s is extra information the method does not need. The survey of 108 is the other: a large sample, yet still a t, because s is an estimate.

The words to listen for are the problem's own. Known, has been consistently, or the population standard deviation is all point to z. A sample standard deviation of, or numbers you must summarise yourself, point to t. If the problem lists raw data and says nothing about sigma, the answer is always t, because the only spread you can have is the one you compute.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution Practice §8.1-8.2, pp. 434-437 — Exercises 13, 23-29, 38-46

14. Trap: a z because the sample is big

Trap

The trap

108 Americans watch a mean 151 hours of TV a month, s = 32. For 99 percent, a first attempt says n is over 30, so:

\[ EBM = 2.576 \cdot \frac{32}{\sqrt{108}} \]

\[ = 2.576(3.079) = 7.93 \]

Wrong distribution, though only slightly wrong number.

The reasoning sounds sensible: a big sample makes s close to sigma. But close is not equal, and the rule is about where the number came from.

The fix

Sigma is not known; 32 came from the survey. The book's rule is a t whenever s stands in for sigma, here with 107 df:

\[ t_{0.005,\,107} = 2.623 \]

\[ EBM = 2.623(3.079) = 8.08 \]

The two answers differ by only 2 percent here, which is why the old rule survived. But the method should not depend on whether the error happens to be small. Read where s came from, not how big n is.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution Practice §8.2, p. 437 — Exercises 43-47

15. The t distribution and degrees of freedom

Section

Part 2

16. Properties of Student's t distribution

Concept

The book lists the properties you need, and every one of them can be seen in a picture. Here are three t curves beside the standard normal.

Figure (svg): Four bell-shaped curves centred at zero on the same axes, t from minus 4 to 4. The lowest peak and fattest tails belong to t with 1 degree of freedom, peak 0.318. Then t with 3 degrees of freedom, peak 0.368, then t with 10, peak 0.389. The standard normal, dashed, is tallest at 0.399 and thinnest in the tails.

Fewer degrees of freedom: a lower centre and heavier tails. More: the curve climbs towards the dashed normal and its tails thin out.

Read the peaks off the figure. With one degree of freedom the centre is only 0.318 high; with ten it is 0.389, already close to the normal's 0.399. The area under every curve is one, so whatever height the centre gives up must reappear in the tails, and the red curve's tails are visibly fat even at 4.

The last property is different in kind: it is not about the curve but about when you may use it. The t curve is exact only if the individual observations come from an approximately normal population and the sample is random. Part 5 is about checking both. For now, notice that every other property follows from one fact: s wobbles, and the wobble shrinks as df grows.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — Properties of Student's t-Distribution

17. Degrees of freedom: why n minus 1

Concept

The degrees of freedom come from the way s is computed. To find s you take the n deviations of the data from their own mean, square them, add them, and divide by n minus 1.

\[ s = \sqrt{\frac{\sum (x - \bar{x})^2}{n - 1}} \]

Those n deviations are not free to be anything. Because the mean is the balance point of the data, the deviations always add up to zero:

\[ \sum (x - \bar{x}) = \sum x - n\bar{x} = n\bar{x} - n\bar{x} = 0 \]

The middle step uses the definition of the mean: the sum of the data is n times the mean. So once you know any n minus 1 of the deviations, the last one is forced: it is whatever brings the total back to zero.

The book puts it this way: the other n minus 1 deviations can change or vary freely, and we call the number n minus 1 the degrees of freedom. A sample of 15 has 14 degrees of freedom; a sample of 20 has 19. One piece of information was spent estimating the mean before s could be computed.

\[ df = n - 1 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — the degrees of freedom

18. Worked: the forced deviation in a sample of four

Worked example

Take the first four acupuncture readings, 8.6, 9.4, 7.9 and 6.8. Compute their deviations from their mean, check they sum to zero, and show that three of them determine the fourth.

The plan is to compute the mean, then three deviations, and then find the fourth two ways: once directly from the data, and once from the rule that the total must be zero. If the rule is real, the two agree.

Add the data

Why: four values.

\[ 8.6 + 9.4 + 7.9 + 6.8 = 32.7 \]

Divide by n for the mean

Why: the balance point.

\[ \bar{x} = \frac{32.7}{4} = 8.175 \]

The first three deviations

Why: each value minus the mean.

\[ 0.425, \quad 1.225, \quad -0.275 \]

Their total

Why: these three are free.

\[ 0.425 + 1.225 - 0.275 = 1.375 \]

The fourth is forced

Why: the four must add to zero.

\[ d_4 = 0 - 1.375 = -1.375 \]

Figure (svg): A number line from 6.5 to 9.5 with four dots at 6.8, 7.9, 8.6 and 9.4, and a dashed yellow line at their mean, 8.175. From the mean an arrow runs to each value: plus 0.425 and plus 1.225 to the right, minus 0.275 and minus 1.375 to the left. The last arrow, minus 1.375, is pink and marked forced.

The arrows to the right and the arrows to the left have the same total length. Once three are drawn, the fourth has no choice.

So with four values there were only three free deviations: three degrees of freedom. The sum of squares that builds s is made from four squared deviations, but only three of them carry independent information, which is why s squared divides by three and why the t-score from this sample follows the t distribution with 3 df.

Check against the data directly

Why: the fourth value minus the mean must give the forced deviation.

\[ 6.8 - 8.175 = -1.375 \; \checkmark \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — the deviations sum to zero

19. Degrees of freedom and the notation

Fill the middle

The book writes the distribution of the t-score as T following t sub df, with df equal to n minus 1. For its own example, a sample of 20 items gives df equal to 19, written T following t sub 19. Fill in the degrees of freedom for this section's other samples.

Fill in the blanks

n = 15: \; T \sim t_14 \qquad n = 12: \; T \sim t_11 \qquad n = 70: \; df = 69

Why: Subtract one from the sample size every time: the acupuncture sample of 15 has 14 degrees of freedom, the hypnotherapy sample of 12 has 11, and the 70 ER patients have 69. The subscript names which t curve to use.

The subscript is not decoration. Each df is a different curve with a different multiplier, so writing the subscript is how you remember which row of the table, or which input to the calculator, you need.

A slip worth catching now: df is about the sample size, not the confidence level and not the population size. A 95 percent interval from 15 people and a 99 percent interval from 15 people both use 14 df; only the tail area changes.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — T ~ t_df, df = n − 1, and the example T ~ t_19

20. Slide the degrees of freedom

Tweak it

Here is the t density with df on a slider. The standard normal's peak height is 0.399; keep that number in mind as you move the slider.

Parameter explorer

Slide df from 1 to 30. Read the height of the peak at t equal to 0, and the height of the curve at t equal to 3. Which way does each move as df grows, and roughly where does the peak stop changing much?

\[ T \sim t_{{df}}, \qquad \text{normal peak} = 0.399 \]

  • df — from 1 to 30: degrees of freedom

What to notice: the peak rises and the tails fall at the same time, because the total area is fixed at one. At df equal to 1 the curve's height at 3 is about seven times the normal's; by df equal to 30 the two are hard to tell apart by eye.

This is the book's last property in motion: as the degrees of freedom increase, the graph of Student's t distribution becomes more like the graph of the standard normal. A bigger sample gives a better s, so less extra uncertainty to pay for.

21. Finding the multiplier: t sub alpha over 2

Concept

For an interval you need the t value that cuts off a given tail. Write the confidence level as CL; the area left outside is alpha, one minus CL, split equally between the two tails. The book's t sub alpha over 2 is the t-score with area alpha over 2 to its right.

\[ \alpha = 1 - CL, \qquad t_{\alpha/2} : \text{ right-tail area } \frac{\alpha}{2}, \; df = n - 1 \]

For a 95 percent interval from 15 observations: alpha is 0.05, each tail gets 0.025, and the curve is t with 14 df.

Figure (svg): The t curve with 14 degrees of freedom from minus 4 to 4. The middle region between minus 2.145 and 2.145 is shaded blue and labelled 0.95; each tail beyond those points is shaded yellow and labelled 0.025.

The book's t sub 0.025 is the point with 0.025 to its right. By symmetry the same distance on the left cuts off another 0.025, leaving 0.95 between.

The book gives three ways to get the number. A TI-84 has invT, which takes the area to the LEFT, so you enter one minus 0.025, which is 0.975, and the df. A printed t table (Appendix H) gives t-scores by df row and right-tail area or confidence-level column. Or any statistics software.

\[ \text{invT}(0.975,\; 14) = 2.1448 \]

Notice the input to invT. Entering 0.025 would return minus 2.1448, the matching point on the left; entering 0.95 would return the multiplier for a 90 percent interval. Sketching the curve with its two tails, as here, makes the right input obvious.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 417-418 — invT(area to the left, df) and the t table

22. Worked: the two multipliers the book's examples need

Worked example

Example 8.8 needs a 95 percent interval from 15 observations, and Example 8.9 a 90 percent interval from 20. Find both t multipliers, and confirm each by the tail area it should cut off.

Each lookup has the same three moves: degrees of freedom from n, the tail area from the confidence level, and then the left area that invT wants. Doing them in that order avoids the two common slips, using n as df and putting the whole of alpha in one tail.

Example 8.8: degrees of freedom

Why: 15 subjects.

\[ df = 15 - 1 = 14 \]

Example 8.8: tail area

Why: 95 percent leaves 0.05, split in two.

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

Example 8.8: invert

Why: the area to the left is 0.975.

\[ t_{0.025} = \text{invT}(0.975,\;14) = 2.1448 \]

Example 8.9: degrees of freedom

Why: 20 infants.

\[ df = 20 - 1 = 19 \]

Example 8.9: tail area

Why: 90 percent leaves 0.10, split in two.

\[ \alpha = 0.10, \quad \frac{\alpha}{2} = 0.05 \]

Example 8.9: invert

Why: the area to the left is 0.95.

\[ t_{0.05} = \text{invT}(0.95,\;19) = 1.729 \]

Figure (svg): The t curve with 19 degrees of freedom, with the middle region between minus 1.729 and 1.729 shaded blue and labelled 0.90, and each tail shaded yellow and labelled 0.05. Short dashed orange marks at plus and minus 1.645 show where the normal would cut.

Ninety percent confidence leaves 0.10 outside, split 0.05 and 0.05. The t point, 1.729, sits a little further out than the normal's 1.645.

Compare each with the normal: 2.145 against 1.96, and 1.729 against 1.645. Both t values are larger, and the gap is bigger for the smaller sample, as the slider showed.

Check with tcdf

Why: the area to the right of each multiplier must be the tail you asked for.

\[ \text{tcdf}(2.1448,\;\infty,\;14) = 0.0250, \quad \text{tcdf}(1.729,\;\infty,\;19) = 0.0500 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 418-419 — Examples 8.8 and 8.9

23. Probabilities from a t curve: tcdf

Concept

An interval needs the inverse question, a t value from an area. But the book also gives the forward tool, and it is worth one slide, because it is how you check any multiplier and how Chapter 9 will use the t curve. On a TI-83 or TI-84 the command is tcdf, with lower bound, upper bound and degrees of freedom.

\[ \text{tcdf}(\text{lower},\; \text{upper},\; df) = P(\text{lower} < T < \text{upper}) \]

Go back to the t-score of 0.525 from Part 1, the acupuncture mean measured against a claimed population mean of 8. How unusual is a t-score that large or larger, on 14 degrees of freedom?

\[ P(T > 0.525) = \text{tcdf}(0.525,\; 10^{99},\; 14) = 0.304 \]

About 30 percent of samples would land at least that far above the claimed mean, so the sample is no evidence against it. Use a very large number such as ten to the 99 for an upper bound of infinity, as the calculator requires.

Now compare a familiar normal fact. Within one standard deviation of the mean, the normal holds 0.6827. The t curve with 14 df holds a little less, because more of its area sits out in the tails:

\[ P(-1 < T < 1) = \text{tcdf}(-1,\; 1,\; 14) = 0.666 \]

The difference is small at 14 df and would be large at 2. Either way, the habit is the same: before trusting a t probability, picture the curve and shade the region, so you know whether the answer should be bigger or smaller than one half.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — tcdf(lower bound, upper bound, degrees of freedom)

24. How much bigger than 1.96?

Estimation

A sample of only five observations, 95 percent confidence. Before you look anything up, estimate the multiplier.

Predict first

Roughly what is the 95 percent t multiplier with 4 degrees of freedom?

  • About 2.0, barely above 1.96
  • About 2.8
  • About 4.3
  • About 1.6, below 1.96

Correct: About 2.8

Why: With 4 degrees of freedom the point with 0.025 to its right is 2.776. That is 42 percent larger than 1.96, so an interval from five observations is 42 percent wider than it would be if sigma were known and equal to s.

\[ t_{0.025,\,4} = 2.776, \qquad \frac{2.776}{1.960} = 1.42 \]

The other values in the list are not random. 4.3 is the multiplier for a sample of three, two degrees of freedom: the penalty grows explosively as the sample shrinks. And no t multiplier is ever below the normal's, because the t curve always has heavier tails, so it always has to go further out to leave the same area beyond.

Keep 2.78 for five and about 2.1 for fifteen as rough anchors. They let you sanity-check any multiplier you look up.

Why is the jump so large for a sample of five? With four degrees of freedom, s rests on only four free deviations, so it is a shaky estimate, and the simulation in Part 1 showed how often a small sample's s comes out far too small. The multiplier grows to cover exactly that risk.

25. Order the multipliers

Ranking

Five multipliers that will come up in this deck. Put them in order from smallest to largest, reasoning from two facts: a higher confidence level needs a point further out, and fewer degrees of freedom need a point further out.

Put in order

Order these critical values from smallest to largest.

  1. z for 95 percent
  2. t for 95 percent with 69 df
  3. t for 95 percent with 14 df
  4. t for 95 percent with 4 df
  5. t for 99 percent with 14 df

Why: The z value 1.960 is the floor. With 69 df the t value is 1.995, barely above it; with 14 df it is 2.145; with 4 df it is 2.776. Raising the confidence to 99 percent with 14 df moves the point out to 2.977, further than any of the 95 percent values.

\[ 1.960 < 1.995 < 2.145 < 2.776 < 2.977 \]

Two separate dials move the multiplier. The confidence level decides which tail area you need; the degrees of freedom decide which curve you cut it from. The 99 percent value with 14 df lands beyond even the 95 percent value with only 4 df, so the confidence-level dial can outweigh the sample-size dial.

The smallest one, 1.995 on 69 df, is worth a second look. It sits barely above the normal's 1.96, which is why the ER exercise in Part 3 gives almost the same interval either way. Being correct about the curve matters most when the sample is small, which is exactly when it is easiest to forget.

26. Building the interval

Section

Part 3

27. The same shape, with s and t

Concept

The interval keeps the shape from Section 8.1: a point estimate plus and minus an error bound. Only the error bound changes, in two places. Here is the book's formula when the population standard deviation is unknown:

\[ EBM = t_{\alpha/2}\left(\frac{s}{\sqrt{n}}\right) \]

Here t sub alpha over 2 is the t-score with area alpha over 2 to its right, from the curve with n minus 1 degrees of freedom, and s is the sample standard deviation. Compare it with Section 8.1, where the multiplier was z and the spread was sigma.

\[ (\bar{x} - EBM,\; \bar{x} + EBM) \]

So the procedure for a t interval is five steps. Compute the sample mean and s. Find df, n minus 1. Find alpha over 2 from the confidence level. Look up t sub alpha over 2. Multiply by s over root n for the error bound, then go that far either side of the mean.

A TI-83 or TI-84 does all of this in one command, TInterval under STAT, TESTS. Give it the data list or the summary statistics and the confidence level. Use it to check, but work the steps by hand until each one makes sense, because the steps are where the reasoning is.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 417-418 — the error bound, the interval, and TInterval

28. Reading the error bound piece by piece

Notation

Each piece of the error bound has a job, and each is a place where a mistake can enter. Hover over each before the worked examples.

Annotate

On: \( EBM = t_{\alpha/2}\left(\frac{s}{\sqrt{n}}\right), \quad df = n - 1 \)

  • Area α/2 to its RIGHT, not α. For 95 percent that is 0.025 in each tail.
  • Picks which t curve. A different sample size means a different multiplier.
  • The sample standard deviation, computed from the data with the n − 1 divisor.
  • Divides s down to the spread of sample MEANS. Forgetting it gives an interval as wide as the data.

Read it as a sentence: go out as many estimated standard errors as the t curve requires for your confidence level. The bracket, s over root n, is the estimated standard error; t says how many of them.

Three things make the bound bigger: a higher confidence level, through t; a more spread-out sample, through s; and a smaller sample, which hits twice, through root n and through a larger t. That double hit is the price of small samples, and the next examples show its size.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 417-418 — EBM and its definitions

29. Example 8.8, part 1: the mean and s of fifteen sensory rates

Worked example

A study of acupuncture measures sensory rates for 15 subjects: 8.6, 9.4, 7.9, 6.8, 8.3, 7.3, 9.2, 9.6, 8.7, 11.4, 10.3, 5.4, 8.1, 5.5, 6.9. Before the interval, you need the sample mean and the sample standard deviation.

The book gives 8.2267 and 1.6722 without the arithmetic. Here is the arithmetic, using the shortcut for the sum of squared deviations, which saves writing fifteen subtractions.

Add the fifteen rates

Why: the first ingredient of the mean.

\[ \sum x = 123.4 \]

Divide by n

Why: the sample mean.

\[ \bar{x} = \frac{123.4}{15} = 8.2267 \]

Add the squares of the rates

Why: the second ingredient of s.

\[ \sum x^2 = 1054.32 \]

Sum of squared deviations, by the shortcut

Why: the sum of squares minus the squared sum over n.

\[ \sum (x - \bar{x})^2 = 1054.32 - \frac{123.4^2}{15} \]

Evaluate the correction term

Why: 123.4 squared is 15227.56.

\[ = 1054.32 - 1015.1707 = 39.1493 \]

Divide by the degrees of freedom

Why: n minus 1, which is 14.

\[ s^2 = \frac{39.1493}{14} = 2.7964 \]

Take the square root

Why: back to the units of the data.

\[ s = \sqrt{2.7964} = 1.6722 \]

The divisor in the second-to-last line is the degrees of freedom from Part 2. Dividing by 15 instead would give 1.6155, slightly too small, which is exactly the kind of small s that the t distribution is there to guard against.

Check the sum of squares directly

Why: add the fifteen squared deviations from 8.2267 one by one; the largest is 10.07, for 11.4, which is 3.17 above the mean.

\[ 0.139 + 1.377 + \cdots + 1.760 = 39.149 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 418 — Example 8.8: x̄ = 8.2267, s = 1.6722, n = 15

30. Example 8.8, part 2: the 95 percent interval for acupuncture

Worked example

With the sample mean 8.2267, s equal to 1.6722 and n equal to 15, construct a 95 percent confidence interval for the mean sensory rate of the population, which is assumed normal.

Sigma is not given; s was computed from the data. So this is a t interval with 14 degrees of freedom, and the multiplier is the 2.1448 found in Part 2.

Degrees of freedom and tail area

Why: 15 subjects, 95 percent.

\[ df = 14, \quad \frac{\alpha}{2} = 0.025 \]

The multiplier

Why: invT of 0.975 with 14 df.

\[ t_{0.025} = 2.1448 \]

The estimated standard error

Why: s over the square root of 15, which is 3.8730.

\[ \frac{s}{\sqrt{n}} = \frac{1.6722}{3.8730} = 0.4318 \]

The error bound

Why: the multiplier times the standard error.

\[ EBM = 2.1448(0.4318) = 0.9261 \]

Lower bound

Why: subtract from the mean.

\[ 8.2267 - 0.9261 = 7.3006 \]

Upper bound

Why: add to the mean.

\[ 8.2267 + 0.9261 = 9.1528 \]

Figure (svg): A dot plot of the 15 acupuncture sensory rates from 5 to 12. Above it, a blue bar runs from 7.30 to 9.15 with a yellow tick at the sample mean, 8.23. Only 5 of the 15 dots, coloured blue, lie under the bar; the other ten, orange, lie outside it.

The interval is far narrower than the data. It brackets the population mean, not the individual subjects, most of whom fall outside it.

In context: we estimate with 95 percent confidence that the true population mean sensory rate is between 7.30 and 9.15. The book prints the error bound as 0.9240, a misprint: 2.1448 times 0.4318 is 0.9261, and only 0.9261 reproduces the book's own endpoints and the calculator's (7.3006, 9.1527).

Check against TInterval

Why: the calculator returns (7.3006, 9.1527); the last digit differs from the hand result only because the hand result carries rounded inputs.

\[ \frac{7.3006 + 9.1527}{2} = 8.2267, \quad \frac{9.1527 - 7.3006}{2} = 0.9261 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 418-419 — Example 8.8

31. Try It 8.8: hours of sleep under hypnotherapy

Worked example

A study of hypnotherapy measures hours of sleep for 12 subjects: 8.2, 9.1, 7.7, 8.6, 6.9, 11.2, 10.1, 9.9, 8.9, 9.2, 7.5, 10.5. Construct a 95 percent confidence interval for the mean hours slept in the population, assumed normal.

Same procedure as the acupuncture study, now done start to finish from raw data. No sigma is given, so it is a t interval with 11 degrees of freedom.

Sum and mean

Why: twelve values.

\[ \sum x = 107.8, \quad \bar{x} = \frac{107.8}{12} = 8.9833 \]

Sum of squared deviations

Why: the shortcut, with the sum of squares 986.72.

\[ 986.72 - \frac{107.8^2}{12} = 986.72 - 968.4033 = 18.3167 \]

The standard deviation

Why: divide by 11, then take the root.

\[ s = \sqrt{\frac{18.3167}{11}} = \sqrt{1.6652} = 1.2904 \]

The multiplier

Why: df is 11, and 0.975 to the left.

\[ t_{0.025,\,11} = 2.2010 \]

The error bound

Why: the standard error is 1.2904 over the root of 12, 0.3725.

\[ EBM = 2.2010(0.3725) = 0.8199 \]

The interval

Why: the mean plus and minus the bound.

\[ (8.9833 - 0.8199,\; 8.9833 + 0.8199) = (8.1634,\; 9.8032) \]

Figure (svg): A dot plot of the 12 hours-of-sleep values from 6.9 to 11.2. Above it a blue bar runs from 8.16 to 9.80 hours, with a yellow tick at the sample mean, 8.98.

Twelve values, roughly symmetric with no outliers: a picture consistent with the assumption of a normal population.

In context: we estimate with 95 percent confidence that the mean nightly sleep of the population these subjects represent is between about 8.16 and 9.80 hours. The dot plot is roughly symmetric with no outliers, so the normal assumption is not contradicted by the data.

Check the width

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

\[ 9.8032 - 8.1634 = 1.6398 = 2(0.8199) \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 419 — Try It 8.8

32. Example 8.9: industrial chemicals in newborns' cord blood

Worked example

The Human Toxome Project tested the cord blood of 20 newborn infants for 430 industrial compounds. The numbers of targeted chemicals found were 79, 145, 147, 160, 116, 100, 159, 151, 156, 126, 137, 83, 156, 94, 121, 144, 123, 114, 139, 99. Construct a 90 percent confidence interval for the mean number found in an infant's blood.

Twenty values, no sigma: a t interval with 19 df. The multiplier, 1.729, was found in Part 2; the work here is the summary statistics.

Sum and mean

Why: twenty counts.

\[ \sum x = 2549, \quad \bar{x} = \frac{2549}{20} = 127.45 \]

Sum of squared deviations

Why: the sum of squares is 337679.

\[ 337679 - \frac{2549^2}{20} = 337679 - 324870.05 = 12808.95 \]

The standard deviation

Why: divide by 19, then take the root.

\[ s = \sqrt{\frac{12808.95}{19}} = \sqrt{674.155} = 25.9645 \]

The standard error

Why: the root of 20 is 4.4721.

\[ \frac{25.9645}{4.4721} = 5.8058 \]

The error bound

Why: t with 0.05 to its right and 19 df.

\[ EBM = 1.729(5.8058) = 10.039 \]

The interval

Why: the mean plus and minus the bound.

\[ (127.45 - 10.039,\; 127.45 + 10.039) = (117.411,\; 137.489) \]

Figure (svg): A dot plot of the 20 cord-blood counts from 79 to 160 chemicals. Above it a blue bar runs from 117.4 to 137.5 with a yellow tick at the sample mean, 127.45.

The counts run from 79 to 160, yet the 90 percent interval for the mean is only about 20 chemicals wide: averages vary far less than individuals.

In context: we estimate with 90 percent confidence that the mean number of targeted industrial chemicals in the cord blood of all US newborns is between about 117.4 and 137.5. The book prints 10.038 and (117.412, 137.488); from the data the bound is 10.039, and the calculator's (117.41, 137.49) agrees with this rounding.

Check with TInterval

Why: the centre and half-width of the calculator's interval give back the mean and the bound.

\[ \frac{117.41 + 137.49}{2} = 127.45, \quad \frac{137.49 - 117.41}{2} = 10.04 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 419-420 — Example 8.9 and Table 8.3

33. Which number goes where in Example 8.9?

Matching

A t interval has six ingredients, and the commonest errors are putting the right number in the wrong slot. Match each symbol to its value in the cord-blood example.

Match the pairs

  • a. n
  • b. df
  • c. α/2
  • d. t with α/2
  • e. s/√n
  • p. 20
  • q. 19
  • r. 0.05
  • s. 1.729
  • u. 5.8058

Why: Twenty infants give 19 degrees of freedom. A 90 percent level leaves 0.10 outside, so 0.05 in each tail, and the t value with that area to its right on 19 df is 1.729. The standard error is s, 25.9645, over the root of 20.

Two pairs are easy to cross. The sample size and the degrees of freedom differ by one, and it is the degrees of freedom that choose the curve while the sample size goes under the square root. The tail area and the confidence level are also easy to swap: 0.05 is what lies beyond each cut point, not the confidence.

Once every slot is filled correctly, the error bound is just a product, 1.729 times 5.8058, which is 10.039.

A good habit is to write these five values in a column before computing anything, with their names beside them, exactly as the book's solutions do. Then each one goes into the formula from a labelled place, and a value cannot drift into the wrong slot.

34. Try It 8.9: hours of television, at 98 percent

Worked example

A random sample of statistics students estimated their weekly hours of television: 0, 3, 1, 20, 9, 5, 10, 1, 10, 4, 14, 2, 4, 4, 5. Construct a 98 percent confidence interval for the mean weekly hours.

A new confidence level, so a new tail area: 98 percent leaves 0.02 outside, 0.01 in each tail. Fifteen students give 14 degrees of freedom.

Sum and mean

Why: fifteen values.

\[ \sum x = 92, \quad \bar{x} = \frac{92}{15} = 6.1333 \]

Sum of squared deviations

Why: the sum of squares is 990.

\[ 990 - \frac{92^2}{15} = 990 - 564.2667 = 425.7333 \]

The standard deviation

Why: divide by 14, then take the root.

\[ s = \sqrt{\frac{425.7333}{14}} = \sqrt{30.4095} = 5.5145 \]

The multiplier

Why: 0.01 to the right, so 0.99 to the left, with 14 df.

\[ t_{0.01,\,14} = 2.6245 \]

The error bound

Why: the standard error is 5.5145 over 3.8730, which is 1.4238.

\[ EBM = 2.6245(1.4238) = 3.7368 \]

The interval

Why: the mean plus and minus the bound.

\[ (6.1333 - 3.7368,\; 6.1333 + 3.7368) = (2.3965,\; 9.8702) \]

Figure (svg): A dot plot of the 15 weekly television hours from 0 to 20. Most dots are piled between 0 and 5; a few sit at 9, 10, 10 and 14, and one lies alone at 20. Above it a blue bar runs from 2.40 to 9.87 with a yellow tick at the sample mean, 6.13.

Piled up near zero with a long right tail: this sample does not look like it came from a normal population, which the t interval assumes.

In context: with 98 percent confidence, the mean weekly television time for statistics students is between about 2.4 and 9.9 hours. But look at the dot plot before trusting it: the data pile up near zero with one student at 20 hours. The t interval assumes a normal population, and with only 15 values that skew matters. The step-zero slide in Part 5 comes back to this.

Check the centre

Why: the midpoint of the interval must be the sample mean.

\[ \frac{2.3965 + 9.8702}{2} = 6.1333 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 420 — Try It 8.9 and Table 8.4

35. Trap: s where s over root n belongs

Trap

The trap

For the acupuncture interval, a first attempt multiplies t by s itself:

\[ EBM = 2.1448(1.6722) = 3.587 \]

\[ (4.64,\; 11.81) \]

Wrong. That interval contains every one of the 15 data values.

It looks plausible because s is the number you just worked hardest to compute, so it feels like the thing to use.

The fix

The interval is for the mean, and means vary far less than individuals. Divide s by root n first:

\[ \frac{1.6722}{\sqrt{15}} = 0.4318 \]

\[ EBM = 2.1448(0.4318) = 0.926 \]

The quick check: an interval for a mean should be much narrower than the data's spread. If yours covers nearly all the data, the root n has gone missing.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 418 — Example 8.8

36. Find the error: a 95 percent multiplier that is too small

Error analysis

A student builds Example 8.8's 95 percent interval like this, and gets an interval noticeably narrower than the book's. Find the slip before reading the notes.

Annotate

On: \( EBM = t_{0.05}\left(\frac{1.6722}{\sqrt{15}}\right) = 1.761(0.4318) = 0.760 \)

  • This line puts all of α = 0.05 in the right tail. That is the multiplier for a 90 percent interval.
  • A two-sided interval splits α: 0.025 in each tail, so t with 0.025 = 2.1448 on 14 df.
  • invT(0.95, 14). The right input for 95 percent is invT(0.975, 14).
  • Too small by 18 percent. The interval (7.47, 8.99) would catch μ only about 90 times in 100.

The slip hides because 1.761 is a perfectly reasonable-looking t value. The sketch catches it: draw the curve, shade 95 percent in the middle, and write the area in each tail. With 0.025 on each side, the area to the left of the upper cut point is 0.975, which is what invT needs.

\[ t_{0.025,\,14} = \text{invT}(0.975,\;14) = 2.1448 \]

The consequence is not a small rounding difference. An interval built with the 90 percent multiplier but labelled 95 percent overstates how sure you are: it would miss the true mean twice as often as it claims.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 418 — the area to the right of t sub 0.025 is 0.025

37. Exercises 38-42: emergency-room waits

Worked example

A hospital surveyed 70 randomly chosen ER patients. Their mean wait before being called back was 1.5 hours, with a sample standard deviation of 0.5 hours. Identify the pieces, choose the distribution, and construct a 95 percent confidence interval for the mean wait.

Here you get summary statistics instead of raw data, which skips straight to the interval. The decision still has to be made: 0.5 hours is a sample standard deviation, so this is a t interval, even with 70 patients.

Identify the pieces

Why: Exercise 38's blanks.

\[ \bar{x} = 1.5, \quad s = 0.5, \quad n = 70, \quad n - 1 = 69 \]

Choose the distribution

Why: sigma is unknown; s estimates it.

\[ T \sim t_{69} \]

The multiplier

Why: 0.025 in each tail, 69 df.

\[ t_{0.025,\,69} = 1.9949 \]

The standard error

Why: the root of 70 is 8.3666.

\[ \frac{0.5}{8.3666} = 0.05976 \]

The error bound

Why: multiplier times standard error.

\[ EBM = 1.9949(0.05976) = 0.1192 \]

The interval

Why: the mean plus and minus the bound.

\[ (1.5 - 0.1192,\; 1.5 + 0.1192) = (1.3808,\; 1.6192) \]

In context (Exercise 42): we estimate with 95 percent confidence that the mean time all patients at this hospital wait to be called back is between about 1.38 and 1.62 hours, roughly 83 to 97 minutes. Notice the multiplier, 1.995, is within 2 percent of the normal's 1.96: with 69 df the t correction is small, but it is still the correct curve.

Check the error bound in minutes

Why: a bound of 0.1192 hours is about 7 minutes, and the interval must span twice that around 90 minutes.

\[ 0.1192 \times 60 = 7.15, \quad 90 \pm 7.15 = (82.85,\; 97.15) \text{ min} \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution Practice §8.2, p. 436 — Exercises 38-42

38. Exercises 49-57: colours on national flags

Worked example

A random survey of 39 national flags, chosen with replacement, recorded the number of colours on each: 1 colour on 1 flag, 2 on 7, 3 on 18, 4 on 7, and 5 on 6. Construct a 95 percent confidence interval for the mean number of colours on a national flag.

The data come as a frequency table, so each sum is a sum of value times frequency. No population sigma is known, so this is a t interval with 38 degrees of freedom.

Sum of the values

Why: value times frequency, added.

\[ 1(1) + 2(7) + 3(18) + 4(7) + 5(6) = 127 \]

The sample mean

Why: divide by the 39 flags.

\[ \bar{x} = \frac{127}{39} = 3.2564 \]

Sum of squares

Why: each squared value times its frequency.

\[ 1 + 28 + 162 + 112 + 150 = 453 \]

Sum of squared deviations

Why: the shortcut.

\[ 453 - \frac{127^2}{39} = 453 - 413.5641 = 39.4359 \]

The standard deviation

Why: divide by 38, then take the root.

\[ s = \sqrt{\frac{39.4359}{38}} = \sqrt{1.0378} = 1.0187 \]

The error bound

Why: t is 2.0244 on 38 df; the standard error is 1.0187 over 6.2450, which is 0.1631.

\[ EBM = 2.0244(0.1631) = 0.3302 \]

The interval

Why: the mean plus and minus the bound.

\[ (3.2564 - 0.3302,\; 3.2564 + 0.3302) = (2.9262,\; 3.5866) \]

Figure (svg): A bar chart of the number of colours on 39 national flags: 1 flag with one colour, 7 with two, 18 with three, 7 with four and 6 with five. A dashed yellow line marks the mean, 3.26, and a blue band from 2.93 to 3.59 shows the 95 percent interval.

The data are whole numbers from 1 to 5, not a bell. With 39 flags the sample mean is still close to normal, so the interval is trustworthy.

In context: we estimate with 95 percent confidence that the mean number of colours on national flags is between about 2.93 and 3.59. The individual counts are whole numbers and not remotely normal, but with 39 flags the sample mean is close to normal by the central limit theorem, which is what the method needs.

Check the mean against the bar chart

Why: the balance point of bars at 1 to 5 with heights 1, 7, 18, 7, 6 must sit a little above 3, since the right side carries 13 flags and the left only 8.

\[ 3.2564 - 3 = 0.2564 > 0 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution Practice §8.2, pp. 437-438 — Exercises 49-57 and Table 8.10

39. Working backwards: recovering s from a published interval

Worked example

A report gives only a 95 percent t interval, (7.30, 9.15), from 15 subjects. Section 8.1 showed how to recover the mean and error bound from an interval. Go one step further and recover the sample standard deviation.

The interval is centred on the mean and its half-width is the error bound. The error bound is t times s over root n, and t and n are known, so s is the only unknown left. Solve for it.

The sample mean is the midpoint

Why: average the two endpoints.

\[ \bar{x} = \frac{7.30 + 9.15}{2} = 8.225 \]

The error bound is the half-width

Why: half the difference.

\[ EBM = \frac{9.15 - 7.30}{2} = 0.925 \]

Write the error bound formula

Why: the only unknown is s.

\[ 0.925 = 2.1448\left(\frac{s}{\sqrt{15}}\right) \]

Multiply both sides by root 15

Why: the root of 15 is 3.8730.

\[ 0.925(3.8730) = 2.1448\,s \;\Rightarrow\; 3.5825 = 2.1448\,s \]

Divide by the multiplier

Why: s on its own.

\[ s = \frac{3.5825}{2.1448} = 1.670 \]

That is the acupuncture study again, and the recovered 1.670 differs from the true 1.6722 only because the endpoints were rounded to two decimals. Working backwards like this lets you check a published interval: if the reported s and n do not reproduce it, something in the report is wrong.

Check against the original data

Why: the acupuncture sample had s equal to 1.6722; the difference comes from rounding the endpoints.

\[ \frac{1.670}{1.6722} = 0.9987 \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.1, pp. 415-416 — working backwards; §8.2, p. 418, Example 8.8

40. How t compares with z

Section

Part 4

41. Always wider, and converging to z

Concept

Put the t multiplier and the z multiplier side by side for the same confidence level. Every t value is larger, because the t curve's heavier tails make it go further out to leave the same area beyond. And every t value approaches z as the sample grows, because a large sample's s is a very good estimate of sigma.

Figure (svg): Dots for the 95 percent t multiplier at sample sizes 3 to 40. At n equals 3 it is 4.30; at 5 it is 2.78; at 10 it is 2.26; at 15 it is 2.14; by 30 it is 2.05 and by 40 it is 2.02. A dashed orange line marks the normal's 1.96, which the dots approach from above and never reach.

The penalty for estimating sigma collapses fast. Most of the drop happens before n reaches ten; after thirty the dots hug the normal's line.
ndft for 95%t ÷ 1.96extra width
324.3032.195120%
542.7761.41742%
15142.1451.0949.4%
30292.0451.0444.4%
70691.9951.0181.8%
10009991.9621.0010.12%

The last column is how much wider a t interval is than a z interval that used the same s as if it were sigma. Read it as the price of not knowing sigma. With three observations the price more than doubles the interval; with fifteen it is under a tenth; with a thousand it is invisible.

So the old n over 30 rule was not crazy: past thirty the price is a few percent. But it was never zero, and today nothing is gained by ignoring it.

42. The price of not knowing sigma, drawn

Picture it

The table gave the price as a percentage. Here it is as a picture, for the acupuncture sample's 14 degrees of freedom. The solid curve is the t distribution, the dashed one the normal that Section 8.1 would use.

Figure (svg): The t curve with 14 degrees of freedom, solid blue, and the standard normal, dashed orange, on the same axes from minus 4 to 4. On the right, the region under the t curve beyond 1.96 is shaded red and labelled 0.035; the further region beyond 2.145 is shaded yellow and labelled 0.025. Arrows mark the gap between 1.96 and 2.145 on the axis.

Cut the t curve at the normal's 1.96 and 0.035 is left beyond, not 0.025. To leave only 0.025 you must walk out to 2.145; that walk is the extra width.

Start at the normal's 95 percent cut point, 1.96. Under the normal, 0.025 lies beyond it. Under the t curve, with its heavier tail, 0.035 lies beyond it: both the red sliver and the yellow tail. Using 1.96 with s would therefore leave about 0.07 outside the interval in total, not 0.05.

\[ P(T_{14} > 1.96) = 0.035, \qquad P(T_{14} > 2.145) = 0.025 \]

To get back down to 0.025 in the tail you have to walk further out along the axis, until the heavier tail has thinned to the same area. That happens at 2.145. The red sliver, between 1.96 and 2.145, is the probability you would be ignoring, and the distance across it is the extra width you pay for.

As the degrees of freedom grow, the solid curve sinks onto the dashed one, the red sliver shrinks, and the two cut points merge. That is the convergence in the table, seen in the tail where it matters.

43. Worked: the acupuncture data both ways

Worked example

Use the acupuncture summary, mean 8.2267, s equal to 1.6722, n equal to 15. Build a 95 percent interval two ways, first wrongly treating s as if it were sigma with 1.96, then correctly with t, and measure the difference.

The point of doing the wrong version is to see exactly what the t distribution adds. Everything except the multiplier is shared, so the difference in width is the price of estimating sigma, and nothing else.

The shared standard error

Why: same s, same n.

\[ \frac{1.6722}{\sqrt{15}} = 0.4318 \]

The z version's bound

Why: 1.96 times the standard error.

\[ 1.96(0.4318) = 0.8463 \]

The z version's interval

Why: the mean plus and minus 0.8463.

\[ (7.3804,\; 9.0730) \]

The t version's bound

Why: 2.1448 times the standard error.

\[ 2.1448(0.4318) = 0.9261 \]

The t version's interval

Why: Example 8.8's answer.

\[ (7.3006,\; 9.1528) \]

Figure (svg): Two horizontal interval bars over a number line from 7 to 9.5, both centred on the dashed yellow sample mean 8.23. The upper, orange bar, using 1.96 with s, runs from 7.38 to 9.07. The lower, blue bar, using the t multiplier 2.145, runs from 7.30 to 9.15 and is visibly longer at both ends.

Same centre, same s, same n. Only the multiplier differs, so the t interval is longer by the ratio 2.145 to 1.96, about 9 percent.

In context: the honest interval reaches about 0.08 further at each end. That extra width is not a penalty for a mistake; it is the uncertainty in s itself, which the z version silently ignored. A z interval built this way would really capture the true mean less often than 95 percent of the time.

Check the ratio of widths

Why: it must equal the ratio of the multipliers, since everything else is shared.

\[ \frac{0.9261}{0.8463} = 1.094 = \frac{2.1448}{1.96} \]

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 417-418 — Example 8.8

44. Worked: how much wider at five observations, and at a thousand?

Worked example

Two studies use 95 percent intervals with s standing in for sigma: one has 5 observations, the other 1000. For each, how much wider is the t interval than a z interval built with the same s?

The standard error cancels when you compare the two intervals from the same study, so the whole answer is in the ratio of the multipliers. Find each t, divide by 1.96, and read the excess as a percentage.

Small study: degrees of freedom

Why: five observations.

\[ df = 5 - 1 = 4 \]

Small study: multiplier

Why: 0.975 to the left, 4 df.

\[ t_{0.025,\,4} = 2.7764 \]

Small study: ratio

Why: t over z.

\[ \frac{2.7764}{1.9600} = 1.4166 \;\Rightarrow\; 42\% \text{ wider} \]

Large study: degrees of freedom

Why: a thousand observations.

\[ df = 1000 - 1 = 999 \]

Large study: multiplier

Why: 0.975 to the left, 999 df.

\[ t_{0.025,\,999} = 1.9623 \]

Large study: ratio

Why: t over z.

\[ \frac{1.9623}{1.9600} = 1.0012 \;\Rightarrow\; 0.12\% \text{ wider} \]

In context: with five observations, not knowing sigma widens the interval by about two fifths; with a thousand it widens it by about a thousandth. That is the summary the section's description gives: the price of not knowing sigma is large for a small sample and negligible for a large one.

Check with the tail areas

Why: the z multiplier, put into the t curve with 4 df, should leave far more than 0.025 beyond it.

\[ P(T_4 > 1.96) = 0.061, \quad P(T_{999} > 1.96) = 0.0251 \]

45. Push the sample down to two

Edge cases

The table stopped at three observations. Push to the smallest sample that has a standard deviation at all: two observations, one degree of freedom.

Discussion prompt

With n equal to 2, find the 95 percent t multiplier and say what it does to the interval. Two observations of a nightly sleep time are 7.5 and 8.5 hours; build the interval.

This is the boundary of the method. With one degree of freedom, s rests on a single deviation, and a pair of values that happen to sit close together would give a tiny s and a wildly overconfident z interval. The t curve with 1 df, the red one with the fat tails in Part 2, protects you by stretching the multiplier to 12.7.

And with n equal to 1 there is no s at all, no degrees of freedom, and no interval. One observation tells you nothing about spread. That is the zero the degrees-of-freedom formula hits when n minus 1 reaches zero.

46. Is the t interval always wider than the z interval?

Counterexample

Part 4 has said a t interval is always wider than a z interval. That claim needs a condition, and the lettuce exercises show it. A sample of 20 heads has mean weight 2.2 pounds with s equal to 0.1, and the population standard deviation is known to be 0.2.

Discussion prompt

Build a 90 percent interval's error bound two ways: correctly, with z and the known sigma; and with t and the sample's s. Which is wider, and does this contradict Part 4?

So the precise claim is: for the same s, the t interval is always wider. A t interval can come out narrower than the true z interval when the sample's s is smaller than sigma, and that is exactly the situation the wedge figure in Part 1 showed, samples with an unluckily small s. Here sigma is known, so the correct choice is z, and the narrow t interval would be overconfident.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution Practice §8.1, p. 435 — Exercises 13-18, the lettuce

47. The z interval against the t interval

Comparison

Section 8.1's interval and this section's interval share a shape and differ in three cells. Fill in the blanks.

Comparison matrix

when to use itspread in the boundmultiplier fromdegrees of freedom95% multiplier, n = 15
z interval (8.1)σ knownσstandard normalnone1.960
t interval (8.2)sigma unknownsStudent's tn - 12.145

Read the matrix by columns. The when column is the decision of Part 1: the deciding fact is whether sigma is known, never the sample size. The spread and multiplier columns are the two changes in the formula. The degrees-of-freedom column is the reason the t multiplier depends on n at all.

The last column puts a number on the difference for the acupuncture sample: 2.145 against 1.960, about 9 percent. Everything else in the two procedures, the centre, the root n, the interpretation, is identical.

One more column could be added: the assumptions. Both intervals assume a random sample, and both are exact for a normal population and approximate otherwise. The t interval does not need any more from the population than the z interval did; it only admits honestly that the spread was estimated.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 417-418 — and §8.1 for the z interval

48. Trap: dividing by root of n minus 1

Trap

The trap

Having learned that df is n minus 1, a student builds the cord-blood standard error with it:

\[ \frac{s}{\sqrt{n - 1}} = \frac{25.9645}{\sqrt{19}} \]

\[ = 5.9567 \]

Wrong. The n minus 1 has already been used.

The fix

The n minus 1 appears twice in the method, in two specific places: inside s, as the divisor of the squared deviations, and as the df that picks the t curve. The standard error divides by root n.

\[ \frac{s}{\sqrt{n}} = \frac{25.9645}{\sqrt{20}} = 5.8058 \]

Using root 19 here would make every interval about 2.6 percent too wide for twenty observations, and more for smaller samples. Keep the three places straight: s divides by n minus 1, df is n minus 1, the standard error divides by root n.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 417-419 — EBM and Example 8.9

49. Reading the result

Section

Part 5

50. What the 95 percent belongs to

Concept

The book's interpretation sentence for Example 8.8 is: we estimate with 95 percent confidence that the true population mean sensory rate is between 7.30 and 9.15. Every word of that sentence is doing work. The interval is about the population mean, it is stated in the units of the data, and it names its confidence level.

What the confidence level means is a statement about the method. Here are 25 samples of ten values, each drawn from a normal population whose mean is 50, and the 95 percent t interval built from each.

Figure (svg): Twenty-five horizontal interval bars stacked one above another, each a 95 percent t interval from a different random sample of ten values taken from a normal population whose mean is 50. A vertical dashed yellow line marks 50. 23 bars cross the line and are blue; 2 miss it and are red. The bars have different lengths because each sample has its own s.

The method catches the true mean about 95 times in 100. Any one interval either contains 50 or does not; the 95 percent belongs to the procedure.

In this run 23 of the 25 intervals caught the true mean and 2 missed. Over many runs the method catches the mean about 95 times in 100. Notice the intervals have different lengths as well as different centres: each sample has its own s, which is exactly the extra wobble the t multiplier pays for.

Once you have computed one interval, the true mean is either inside it or not; the 95 percent does not attach to that one interval as a probability. It describes how often intervals built this way succeed. That is why the sentence says confidence rather than probability.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 418 — the interpretation in Example 8.8

51. The two assumptions, and which one a bigger sample relieves

Concept

The book lists the assumptions behind the t interval among the distribution's properties. There are two, and it insists they are separate.

  1. The population of individual observations is approximately normal. If it is bell-shaped, the assumption is met and needs no further discussion.
  2. The sample is random. The book calls this a completely separate assumption from normality.

The book adds that the size of the population is generally not relevant unless it is very small. A sample of 20 from a population of a million is treated the same as a sample of 20 from ten million.

The two assumptions respond to sample size very differently. Normality matters most for small samples; for larger ones the central limit theorem makes the sample mean nearly normal even when the individuals are not, as the flag counts showed. Randomness is never relieved by sample size. A biased sample of 10,000 gives a very narrow interval around the wrong value.

So check the two differently. For normality, look at a dot plot or histogram of the data for strong skew or outliers, especially when n is small. For randomness, look at how the data were collected; no plot of the data can reveal a biased sampling method.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 417 — the underlying population and random sampling

52. Step zero: should Try It 8.9 have used a t interval?

Step zero

Part 3 built a 98 percent t interval from the 15 students' television hours, (2.40, 9.87). The method assumes the population of individual hours is approximately normal. Before trusting the numbers, look at the data again: 0, 1, 1, 2, 3, 4, 4, 4, 5, 5, 9, 10, 10, 14, 20.

Discussion prompt

Is the normality assumption believable for these data, and what should you say about the interval as a result?

This is the habit to build: look at the data before choosing the method. A dot plot takes ten seconds and is the only check on the normality assumption that the calculation itself cannot do. The sleep data passed it; the television data do not.

Notice the method still ran without complaint. TInterval never checks its own assumptions; it will produce an interval from any list of numbers. Judging whether that interval means what it says is your job, not the calculator's.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 417, 420 — the normality assumption and Try It 8.9

53. Which assumption is at risk?

Sorting

For each study, decide which of the two assumptions is the real worry. Ask first how the data were gathered, then what shape the values are likely to have and how many there are.

Sort into buckets

Sort each study by the assumption most at risk.

normality: small sample, skewed values
ten household incomes, randomly sampled from a city; eight repair times for a machine, where most take minutes but a few take hours
randomness: how the sample was chosen
the 200 customers who chose to fill in an online survey; heights of the 30 players on one college basketball team, used for all students
neither looks like a problem
40 randomly chosen bags of rice from a production line, weights roughly bell-shaped
n
Incomes and repair times are strongly right-skewed, and with ten or eight values the central limit theorem cannot rescue the sample mean.
r
Volunteers who chose to respond, and a basketball team standing in for all students, are not random samples of the population of interest. A bigger sample of the same kind would not help.
ok
A random sample of 40 from a roughly bell-shaped population meets both assumptions comfortably.

The two randomness failures are the more dangerous ones, because nothing in the numbers warns you. The basketball players' heights could look beautifully normal and still give an interval for the wrong population. No amount of data fixes a sample drawn from the wrong place.

The two normality cases differ in a way worth noticing too. Incomes and repair times are skewed by their nature: they cannot go below zero but can run very high. With 40 or more values the sample mean would be nearly normal anyway; with eight or ten it is not, and a t interval's stated confidence is unreliable.

54. Trap: the interval as a range for individuals

Trap

The trap

After Example 8.8, a first reading of the interval says:

\[ \text{95\% of subjects: } 7.30 \text{ to } 9.15 \]

Wrong. Only 5 of the 15 subjects in the sample are inside that range.

The slip comes from reading 95 percent as a share of people. It is a success rate of the method, and the interval is about one number, the mean.

The fix

The interval estimates the population mean. Its width comes from the standard error, s over root n, which shrinks as n grows. Individuals spread with s itself:

\[ \frac{s}{\sqrt{n}} = 0.43 \quad \text{vs} \quad s = 1.67 \]

With a larger sample the interval for the mean would close in towards a single value, while the individuals' spread would stay about the same. A range for individuals needs a different tool, not a confidence interval for the mean.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, p. 418 — Example 8.8

55. What does a bigger sample fix?

Prediction

A researcher worries about two things in a small t interval: that the population may not be normal, and that the volunteers in the study may not be typical. She plans to quadruple the sample by recruiting more volunteers the same way.

Predict first

Which worry does quadrupling the sample in the same way address?

  • The normality worry only
  • The volunteer worry only
  • Both worries
  • Neither worry

Correct: The normality worry only

Why: A bigger sample makes the sample mean closer to normal whatever the population's shape, and it shrinks the multiplier and the standard error. But more volunteers recruited the same way carry the same bias; the interval would just be narrower around the wrong value.

\[ \text{4 times the sample:} \quad \frac{s}{\sqrt{4n}} = \frac{1}{2}\cdot\frac{s}{\sqrt{n}} \]

Quadrupling n halves the standard error, and also lowers the t multiplier, so the interval narrows by a bit more than half. That precision is real, but precision is not accuracy. A narrow interval around a biased estimate is a confident wrong answer.

To fix the volunteer worry she must change how the sample is chosen, for instance by drawing names at random from a list of the whole population, not just collect more of the same.

56. A confidence interval is not a safety limit

Real world

Example 8.9 estimated the mean number of industrial chemicals in newborns' cord blood at between 117.4 and 137.5. Suppose a health agency wanted to set a threshold above which an individual infant's result is unusual.

Discussion prompt

Explain why the confidence interval for the mean is the wrong tool for that threshold, and say roughly where the individual results actually fall in this sample.

This confusion is common in news reports and even in professional settings: a narrow confidence interval for an average is read as a narrow range of typical values. The larger the study, the narrower the interval and the more misleading that reading becomes.

The question to ask whenever you see an interval reported is simply: is this about the average, or about individuals? A t interval from this section is always about the average.

Here the numbers make the difference vivid. The standard error, about 5.8, is the spread of averages of twenty infants. The standard deviation, about 26, is the spread of single infants, more than four times larger, and it is the one a threshold for individuals would have to use.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 419-420 — Example 8.9

57. Putting it together

Section

Part 6

58. Pattern: a confidence interval for a mean, in order

Pattern

Every problem in this section ran through the same sequence. The calculation is short; the decisions come first and last.

Use the diagram as a checklist. The first branch, whether sigma is known, is the decision most often made for the wrong reason, the size of the sample. Ask it from where the standard deviation came from.

Figure (svg): A flow diagram. First box: is sigma known? A yes arrow leads to a box: use z, Section 8.1. A no arrow leads through four boxes in a row: compute the sample mean and s, and df equals n minus 1; find t with area alpha over 2 to its right; compute EBM equals t times s over root n; write the interval and interpret it, checking normality and randomness.

One question decides the branch: was sigma handed to you, or did you compute s from the data? After that the t branch is four mechanical steps.
  1. Is sigma known? Given as a population fact: z, Section 8.1. Computed from the sample: t.
  2. Summarise the data. The sample mean and s, with s dividing the squared deviations by n minus 1.
  3. Degrees of freedom. df equals n minus 1; it picks the t curve.
  4. Multiplier. Area alpha over 2 to the right; invT takes the area to the left, one minus alpha over 2.
  5. Error bound. t times s over root n; the interval is the mean plus and minus it.
  6. Interpret and check. A sentence about the population mean, in units, with the level; then normality (look at the data) and randomness (look at the design).

\[ EBM = t_{\alpha/2}\left(\frac{s}{\sqrt{n}}\right), \quad df = n - 1, \quad (\bar{x} - EBM,\; \bar{x} + EBM) \]

Most wrong answers here are correct arithmetic on a wrong ingredient: z where t belongs, all of alpha in one tail, s without the root n, n in place of n minus 1 or the reverse. Each has a one-line check, and the pattern puts each check where it belongs.

59. Check: choosing the multiplier

Check

Find the degrees of freedom and the tail area before you look at the options.

Two separate pieces of information go into a t multiplier: the sample size and the confidence level. Say each one out loud before choosing.

Check your understanding

A sample of 25 battery lifetimes has a sample standard deviation of 4 hours. Which multiplier belongs in a 90 percent confidence interval for the mean lifetime?

  • A. t with 0.05 to its right on 24 df: 1.711 (correct)
  • B. t with 0.10 to its right on 24 df: 1.318
  • C. z with 0.05 to its right: 1.645
  • D. t with 0.05 to its right on 25 df: 1.708

Answer: A

Why: The standard deviation was computed from the sample, so the t distribution applies, with 25 minus 1, which is 24, degrees of freedom. Ninety percent confidence leaves 0.10 outside, split into 0.05 per tail, so the multiplier has 0.05 to its right: 1.711.

Why B tempts people
That puts the whole of the leftover 0.10 into one tail. A two-sided interval splits it, 0.05 per tail.
Why C tempts people
That is the right tail area, but the normal curve. The spread came from the sample, so sigma is unknown and t is needed.
Why D tempts people
That uses the sample size as the degrees of freedom. The deviations sum to zero, so only 24 of them are free.

In context: the interval would be the sample mean plus and minus 1.711 times 4 over 5, which is about 1.37 hours. Each wrong option is a real, common slip: the whole alpha in one tail, the normal in place of t, and the sample size in place of the degrees of freedom.

Notice how close three of the four numbers are. In a small sample, a wrong choice of curve or df changes the answer only a little, which is why the slip survives; the wrong tail area changes it a lot.

60. Check: computing the error bound

Check

Find the standard error first, then multiply.

Write out s over root n as its own line before multiplying. That one habit prevents the commonest error in this section.

Check your understanding

A random sample of 16 commute times has mean 31 minutes and s = 8 minutes. The 95 percent multiplier on 15 df is 2.131. What is the confidence interval for the mean commute?

  • A. (26.74, 35.26) (correct)
  • B. (13.95, 48.05)
  • C. (27.08, 34.92)
  • D. (26.60, 35.40)

Answer: A

Why: The standard error is 8 over the root of 16, which is 2. The error bound is 2.131 times 2, which is 4.262, so the interval runs from 31 minus 4.262 to 31 plus 4.262, which is 26.74 to 35.26 minutes.

Why B tempts people
That multiplies 2.131 by s itself, 8, forgetting to divide by the root of n. The interval would cover individuals, not the mean.
Why C tempts people
That uses the normal's 1.96 in place of the t multiplier. With s estimated from 16 values, the t value is required.
Why D tempts people
That divides 8 by the root of 15 instead of the root of 16. The degrees of freedom choose the curve; the standard error still uses n.

In context: we estimate with 95 percent confidence that the mean commute time in the population sampled is between about 26.7 and 35.3 minutes.

Check the answer's size before trusting it: a 95 percent interval should be roughly four standard errors wide, and four times 2 is 8; the chosen interval is 8.5 wide. The interval built with s itself is 34 wide, and fails that test at once.

The other two wrong intervals are much closer, and that is the lesson in them. Using 1.96 or the root of 15 changes the bound by only a few tenths of a minute, so no size check will catch them; only knowing which ingredient goes where will. Root n goes in the standard error, n minus 1 picks the curve.

61. Check: reading the interval

Check

Decide what the interval is about before you read the options.

Ask of each statement: is it about the population mean, about individuals, or about this particular sample? Only one kind of statement is supported.

Check your understanding

From 15 subjects, a 95 percent t interval for the mean sensory rate is (7.30, 9.15). Which statement is correct?

  • A. We are 95% confident the population mean sensory rate is between 7.30 and 9.15. (correct)
  • B. 95% of people have a sensory rate between 7.30 and 9.15.
  • C. There is a 95% chance the sample mean is between 7.30 and 9.15.
  • D. If the sample size were larger, the population mean would be closer to 8.23.

Answer: A

Why: The interval estimates the population mean, and the 95 percent describes how often the method succeeds over repeated samples. That is exactly what the confidence sentence says, in the units of the study.

Why B tempts people
That treats the interval as a range for individuals. Only 5 of the 15 subjects in the sample even lie inside it.
Why C tempts people
The sample mean is 8.2267, the centre of the interval, with certainty. Confidence is about the unknown population mean.
Why D tempts people
A larger sample changes the interval, not the population mean, which is a fixed number we are trying to estimate.

Each wrong statement misplaces what the interval is about: individuals, the sample mean, or the parameter itself as if it moved. Get the subject of the sentence right and the rest follows.

The statement about a larger sample is the subtle one. More data would change the interval, narrowing it and moving its centre, but the population mean is a fixed property of the population and does not respond to how many people you measure. What moves is your estimate, not the thing estimated.

A good habit is to write the interpretation sentence with the words population mean in it every time, followed by the units. It is very hard to misread an interval when the sentence says what it is about.

62. Explain why a t interval is wider

Explain it to yourself

The fact that the t interval is wider than the z interval, and by less as the sample grows, came up throughout. Explaining it without formulas is a good test of whether the idea has landed.

Discussion prompt

In two or three sentences, explain to someone who knows Section 8.1 why replacing sigma with s calls for a bigger multiplier, and why the extra width nearly disappears for a large sample.

A strong answer names the two sources of uncertainty and says which one shrinks with n. A very strong one adds the numbers: about 42 percent wider at five observations, under 10 percent at fifteen, and about a tenth of a percent at a thousand.

It also helps to name Gosset's problem: with only a handful of barley samples, the second source of chance was large, and ignoring it gave intervals that missed far more often than claimed.

63. Exit ticket: a small study of reaction times

Exit ticket

A psychology class measures the reaction times, in milliseconds, of 8 randomly chosen volunteers from a large university: 251, 274, 239, 262, 288, 245, 270, 255.

Discussion prompt

Construct a 95 percent confidence interval for the mean reaction time, say what it means, and name one assumption you would want to check.

The deciding moves were the ones from the pattern: s was computed, so t; eight values, so 7 df; 95 percent, so 0.025 to the right. If you got a bound of about 11.3, you used 1.96; if you got about 38.5, you multiplied by s instead of s over root 8.

The volunteers are worth a second look: the problem says randomly chosen, which meets the randomness assumption, but in real classes volunteers are rarely random, and that is the assumption no larger sample can fix.

64. Recap

Recap

When the population standard deviation is unknown, use the sample's s in its place and Student's t distribution in place of the normal. That is the whole change, and it costs some width: the honest price of estimating sigma from the same small sample.

ideaformularead it as
whens used as an estimate for σt, whatever the sample size
t-scoret = (x̄ − μ)/(s/√n)standard errors from μ, like z
distributionT ~ t with df = n − 1one curve for each df
shapesymmetric, mean 0, heavy tailsapproaches the normal as df grows
dfn − 1the deviations sum to zero
error boundEBM = t with α/2 · s/√narea α/2 to the right of t
interval(x̄ − EBM, x̄ + EBM)about the population mean
assumptionsnormal population, random samplen helps the first, never the second

The checks that catch most errors: the multiplier must exceed the normal's for the same level; the interval must be much narrower than the data; the interpretation must name the population mean. The Introductory Business Statistics text covers the same interval in its §8.2 with business data, if you want more practice. Section 8.3 builds the matching interval for a proportion.

OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution §8.2, pp. 416-420 — the whole section

OpenStax Introductory Business Statistics 2e, §8.2 A Confidence Interval When the Population Standard Deviation Is Unknown and Small Sample Case §8.2 A Confidence Interval When the Population Standard Deviation Is Unknown and Small Sample Case — the same interval in business contexts

Sources

  1. OpenStax Introductory Statistics 2e, §8.2 A Single Population Mean using the Student t Distribution — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0, pp. 415-421
  2. OpenStax Introductory Business Statistics 2e, §8.2 A Confidence Interval When the Population Standard Deviation Is Unknown and Small Sample Case — 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