Sections 8.1 and 8.2 estimated a population mean. This one changes the parameter rather than the multiplier: when the data are categorical, the quantity of interest is the proportion of the population in one category, and the underlying distribution is the binomial of section 4.3. The point estimate is the sample proportion, written p prime, the number of successes divided by the number of trials, and its error bound is a z-score times the square root of p prime times q prime over n. No population standard deviation appears anywhere, because a binomial's variability is determined once its probability is. The section also introduces the plus-four adjustment, which adds two successes and two failures to the observed counts before proceeding, and gives more accurate intervals when the confidence level is at least ninety percent and the sample at least ten. Solving the error bound for n gives a planning formula, and using one half for the unknown proportion makes it as large as it can be.
Subject: Statistics · 67 slides · symbolic lesson
Open the interactive version of this deck
Title
Statistics · §8.3
When the data are yes or no: estimating what fraction of a whole population is in one category
Objectives
Sections 8.1 and 8.2 estimated a population mean, an average of numbers. Many questions are not about averages at all: what fraction of voters back a candidate, what share of adults own a smartphone. This section builds the confidence interval for that kind of parameter.
The shape of the interval is the one you already know: a point estimate in the middle, the same error bound subtracted and added. Only two ingredients change, the point estimate and the standard error, and the new standard error needs no sigma at all. Seeing why is the heart of the lesson.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 420-427 — the section
Warm-up
This section stands on two earlier results: the binomial distribution's mean and standard deviation from Chapter 4, and the normal curve's 95 percent cut points from Section 8.1. Recall both with a fair coin before anything new arrives.
Discussion prompt
A fair coin is tossed 100 times and X counts the heads. (a) What are the mean and standard deviation of X? (b) If you divide X by 100 to get the proportion of heads, what are the mean and standard deviation of that proportion? (c) Within what distance of one half do about 95 percent of such proportions fall?
Notice what was never needed: a separate standard deviation for the population. Once you know the probability of a head, the spread of the count and of the proportion follow from it. That single fact is why the interval for a proportion looks the way it does.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — dividing the binomial mean and standard deviation by n
Section
Part 1
Concept
During an election year you read headlines such as: the candidate has 40 percent of the vote, within three percentage points. That sentence is a confidence interval. The sample said 40 percent, the margin of error is 3 points, and the pollster is usually 95 percent confident.
Figure (svg): A number line of the proportion of voters favouring a candidate, from 0.30 to 0.50. A blue interval bar runs from 0.37 to 0.43 with the sample proportion 0.40 marked by a yellow dot at its centre; orange arrows show three percentage points on each side.
\[ (0.40 - 0.03,\ 0.40 + 0.03) = (0.37,\ 0.43) \]
So the pollster would be 95 percent confident that the true proportion of all voters favouring the candidate is between 0.37 and 0.43. The parameter is a proportion of a population, not an average of anything.
The book lists other places the same question arises. Investors want the true proportion of stocks that go up each week. Computer sellers want the proportion of households that own a computer. Each is a fraction of a whole population, estimated from a sample, and each gets the same kind of interval.
The procedure, the book says, is similar to the one for a mean, but the formulas are different. This part is about telling the two situations apart; the next part builds the formula.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 420 — the election poll and other proportion questions
Concept
How do you know you are dealing with a proportion problem? The book gives a two-part test. First, the underlying distribution is binomial: each person or item is a trial that ends in success or failure. Second, there is no mention of a mean or an average.
Figure (svg): Left: a five by five grid of 25 circles, one per student; six are filled red for students who smoked in the past week and nineteen are hollow. Right: the counts x equals 6 and n equals 25, the sample proportion p prime equals 6 over 25 equals 0.24, q prime equals 0.76, and a bar of length one split into a red piece of 0.24 and a blue piece of 0.76.
Example 8.12's students show the idea. Each one answers yes or no to having smoked in the past week, so each contributes a category, not a number. Counting the yeses gives x; dividing by the number asked gives the sample proportion.
\[ X \sim B(n,\ p) \quad \Longrightarrow \quad P' = \frac{X}{n} \]
P prime — The random variable for the sample proportion, the number of successes X divided by the number of trials n. Its value in a particular sample, p′ = x/n, is the point estimate of the population proportion p. Some books write it p-hat.
Success just means the category you are counting, here smoked, even when it is nothing to celebrate. The proportion of failures is whatever is left, written q prime, and the two always add to one.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 420-421 — how do you know you are dealing with a proportion problem
Sorting
Eight questions from this chapter and beyond. Sort each one by the parameter it asks about. Look at what a single person or item contributes: a number on a scale, or a yes or a no.
Sort into buckets
Sort each question by the parameter it estimates.
The chemicals question is the one to pause on. It involves counting, so it can look categorical, but each infant contributes a number, how many chemicals, and the question asks for the average of those numbers. Counting within a person makes a measurement; counting people who say yes makes a proportion.
The words fraction, percent, share and proportion are strong clues, and the word average or mean settles it the other way. When the wording is unclear, picture one row of the data: if it reads yes or no, you are in this section.
Notation
Two symbols carry the whole section. Both are computed from the sample, and both are estimates of population values you do not know.
Annotate
On: \( p' = \frac{x}{n}, \qquad q' = 1 - p' \)
Keep these as decimals. A survey reporting 68 percent gives a p prime of 0.68, not 68, and every formula in this section expects the decimal. Convert back to percent only when you write the final sentence.
\[ p' = 0.68, \quad q' = 0.32, \quad p' + q' = 1 \]
The capital letter P prime is the random variable, the proportion before the sample is drawn; the small p prime is its value in your sample. The same distinction between a random sample mean and the number 68 ran through Section 8.1.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — p prime, q prime, x and n
Intuition
The interval needs a sampling distribution for P prime, and you build it from the binomial in three lines. When n is large and p is not close to zero or one, the count X is approximately normal with the binomial's mean and standard deviation.
\[ X \sim N\left(np,\ \sqrt{npq}\right) \]
Now divide the random variable, its mean and its standard deviation by n. Dividing every value of a variable by a constant divides its centre and its spread by that constant.
\[ P' = \frac{X}{n} \sim N\left(\frac{np}{n},\ \frac{\sqrt{npq}}{n}\right) \]
Simplify each piece. The mean is just p. For the standard deviation, write n as the square root of n squared and bring it inside the root.
\[ \frac{\sqrt{npq}}{n} = \sqrt{\frac{npq}{n^{2}}} = \sqrt{\frac{pq}{n}} \]
\[ P' \sim N\left(p,\ \sqrt{\frac{pq}{n}}\right) \]
Read what this says. Sample proportions centre on the true proportion, so p prime is an unbiased estimate of it, and their spread shrinks with the square root of n, exactly like sample means. The difference is that the spread is built from p itself rather than from a separate sigma.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — the normal distribution of proportions
Picture it
The derivation leaned on the normal approximation to the binomial. Here are two exact distributions of P prime, computed from the binomial formula, with the normal curve of the last slide laid over each.
Figure (svg): Two panels. Left: the exact distribution of the sample proportion for 25 trials with success probability 0.24, drawn as separated blue bars at multiples of 0.04, slightly skewed right, with a yellow normal curve of mean 0.24 and standard deviation 0.085 laid over it. Right: the same for 500 trials with probability 0.842, where the bars are so close together they form a smooth bell that the yellow normal curve, mean 0.842 and standard deviation 0.0163, follows almost exactly.
On the right, Example 8.10's situation: 500 people and a proportion near 0.842. The bars are so dense that they merge into a bell, and the normal curve is almost indistinguishable from them. The approximation is excellent.
On the left, Example 8.12's situation: 25 students and a proportion near 0.24. The possible proportions are only the multiples of 0.04, the bars are visibly separate, and the shape leans slightly right because the curve is squeezed against zero. The normal curve is still a fair summary, but a rougher one.
\[ \text{SE} = \sqrt{\frac{(0.24)(0.76)}{25}} = 0.085, \qquad \sqrt{\frac{(0.842)(0.158)}{500}} = 0.0163 \]
The two things that make the approximation good are a large n and a p away from the ends. That is the content of the condition on the next slide, and later the reason for the plus-four adjustment.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — when n is large and p is not close to zero or one
Concept
The book states one condition for using this interval: the number of successes and the number of failures must both be greater than five. In symbols, n times p prime and n times q prime both exceed five, which is the same as saying x and n minus x both exceed five.
\[ np' = x > 5 \quad \text{and} \quad nq' = n - x > 5 \]
| survey | successes x | failures n − x | condition |
|---|---|---|---|
| smartphones, Example 8.10 | 421 | 79 | met |
| voters, Example 8.11 | 300 | 200 | met |
| smokers, Example 8.12 | 6 | 19 | met, just |
| electric vehicles, Example 8.13 | 13 | 37 | met |
| complications, 2 of 40 | 2 | 38 | fails |
Why both counts? The normal curve is symmetric and runs forever in both directions; a proportion is trapped between zero and one. If either count is small, the true proportion is near an end, the distribution is squeezed against that end, and the symmetric normal curve misrepresents it.
The smokers sit right at the edge, with only six successes. That is exactly the kind of small sample where the book recommends the plus-four method of Part 4. The complications survey fails outright, and you will see what goes wrong when the formula is used anyway.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — successes and failures both greater than five
Prediction
A school nurse checks a random sample of 60 student records and finds that 4 students missed a vaccination. She wants a 95 percent interval for the proportion of all students who missed it. Apply the condition before anyone computes.
Predict first
Does the nurse's sample meet the book's condition for the ordinary interval?
Correct: No, because only 4 successes were observed, which is not more than five
Why: Both counts must exceed five. The failures, 56, are fine, but the successes, 4, are not. A sample of 60 is not large enough by itself when the proportion is near zero, because what matters is how many of each kind you saw. The proportion can still be estimated, just not well by this shortcut.
\[ x = 4 \not> 5, \qquad n - x = 56 > 5 \]
The rule of thumb from means, thirty or more, does not carry over. A sample of 60 with a proportion near 0.07 has a sampling distribution squeezed against zero, and the count of successes is what tells you so.
What could the nurse do? A larger sample would bring more successes. Or, since her level is at least 90 percent and her sample at least ten, the plus-four method of Part 4 is designed for exactly this situation.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — the condition on successes and failures
Trap
Example 8.11: 300 of 500 students are registered voters, 90 percent level. A first attempt writes 60 and 40:
\[ 1.645\sqrt{\frac{(60)(40)}{500}} \]
\[ = 1.645 \times 2.191 = 3.60 \]
Wrong. An error bound of 3.60 on a proportion that cannot exceed 1 is impossible on its face.
The formula works in proportions, decimals between zero and one.
\[ 1.645\sqrt{\frac{(0.6)(0.4)}{500}} \]
\[ = 1.645 \times 0.02191 = 0.036 \]
Using percentages multiplies p prime times q prime by ten thousand, so the root and the error bound come out a hundred times too big. Work in decimals throughout, then say 3.6 percentage points in the sentence.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 423 — Example 8.11
Section
Part 2
Notation
The interval has the same shape as for a mean, the point estimate plus and minus an error bound, now called EBP. Read each piece of the error bound before using it.
Annotate
On: \( \text{EBP} = z_{\alpha/2}\sqrt{\frac{p'q'}{n}} \)
\[ \left(p' - \text{EBP},\ \ p' + \text{EBP}\right) \]
Nothing about the z-score is new. You turn the confidence level into alpha, put half of alpha in each tail, and find the z with one minus alpha over two to its left, exactly as before. Only the quantity it multiplies has changed.
The interval is symmetric about p prime, so the midpoint of a published proportion interval is its p prime and half its length is its EBP, just as for a mean.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — the confidence interval and the error bound for a proportion
Concept
For a mean with sigma known, the error bound used the standard error sigma over root n, and sigma had to be handed to you. For a proportion, the book says, the appropriate standard deviation is the square root of p times q over n, and nothing else is needed.
\[ \text{mean: } \frac{\sigma}{\sqrt{n}} \qquad \text{proportion: } \sqrt{\frac{pq}{n}} \]
The reason is that a yes or no variable has no spread of its own to choose. Code yes as 1 and no as 0. Its standard deviation works out to the square root of p times q: once the proportion of ones is known, so is the spread. A mean has no such link, which is why sigma had to be supplied separately.
There is one catch. The formula needs p and q, the population values, and those are exactly what you do not know. So the book replaces them with the sample estimates p prime and q prime.
\[ \sqrt{\frac{pq}{n}} \quad \longrightarrow \quad \sqrt{\frac{p'q'}{n}} \]
This substitution is a small source of error, because the estimated standard deviation is itself a little off whenever p prime is off. With large samples it hardly matters; with small ones it is the very flaw the plus-four method of Part 4 is designed to repair.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — the appropriate standard deviation for a proportion
Worked example
A market research firm surveys 500 randomly selected adults in a large city; 421 say they own a smartphone. Using a 95 percent confidence level, estimate the true proportion of the city's adults who own one. (The book's solution calls X the number with cell phones; the survey asked about smartphones.)
\[ x = 421, \quad n = 500, \quad \text{CL} = 0.95 \]
The point estimate
Why: successes over trials.
\[ p' = \frac{421}{500} = 0.842 \]
The failures' proportion
Why: whatever is not a success.
\[ q' = 1 - 0.842 = 0.158 \]
Check the condition
Why: both counts must exceed five.
\[ x = 421 > 5, \quad n - x = 79 > 5 \]
The condition is met with room to spare: 421 owners and 79 non-owners. So the sampling distribution of the proportion is close to normal, which is what licenses the z-score in the next line. A z-score is only meaningful when the curve it comes from fits.
The z-score
Why: alpha 0.05, so 0.025 per tail and 0.975 to the left: invNorm(0.975, 0, 1).
\[ z_{0.025} = 1.96 \]
The product under the root
Why: multiply, then divide by n.
\[ \frac{p'q'}{n} = \frac{(0.842)(0.158)}{500} = \frac{0.133036}{500} = 0.000266 \]
The standard error
Why: take the square root.
\[ \sqrt{0.000266} = 0.01631 \]
The error bound
Why: z standard errors.
\[ \text{EBP} = (1.96)(0.01631) = 0.0320 \]
The margin is about 3.2 percentage points. Five hundred people pin down a proportion this well because each answer carries little uncertainty when the split is as lopsided as 84 to 16.
The interval
Why: subtract and add.
\[ (0.842 - 0.032,\ 0.842 + 0.032) = (0.810,\ 0.874) \]
Figure (svg): A bell curve centred at the sample proportion 0.842 with standard error 0.0163, its middle 0.95 shaded between 0.810 and 0.874, both marked with yellow lines. Below the axis a blue interval bar with a yellow centre dot at 0.842 runs from 0.810 to 0.874.
In context: we estimate with 95 percent confidence that between 81 percent and 87.4 percent of all adult residents of this city have smartphones. The calculator's 1-PropZInt prints (0.81003, 0.87397); it uses z of 1.95996, which agrees to three decimals.
Check the midpoint and the range
Why: the centre must be p prime, and both ends must lie between zero and one.
\[ \frac{0.810 + 0.874}{2} = 0.842, \quad 0 < 0.810 < 0.874 < 1 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 421-422 — Example 8.10
Fill the middle
Of 250 randomly selected people, 98 report owning a tablet. Build the 95 percent interval for the true proportion of tablet owners. Fill in the point estimate and the error bound, to three decimals.
Fill in the blanks
p' = \frac0.3920.061 = ___, \quad \text___ = 1.96\sqrt______} = ___
Why: Ninety-eight divided by 250 is 0.392, and q prime is 0.608. Their product is 0.2383; divided by 250 it is 0.000953, whose square root is 0.0309. Times 1.96 that is about 0.061.
\[ (0.392 - 0.061,\ 0.392 + 0.061) = (0.331,\ 0.453) \]
Unrounded, the endpoints are 0.3315 and 0.4525, so a calculator shows about (0.332, 0.453). In context: we estimate with 95 percent confidence that between 33.2 and 45.3 percent of all people own a tablet.
Compare with the smartphone survey. Half the sample and a proportion much nearer one half have nearly doubled the error bound, from 0.032 to 0.061. Both changes push the same way; the next slides separate them.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 422-423 — Try It 8.10
Worked example
A political science student surveys 500 students at a large university and finds that 300 are registered voters. Compute a 90 percent confidence interval for the true percent of students who are registered voters, and interpret it.
\[ x = 300, \quad n = 500, \quad \text{CL} = 0.90 \]
The point estimate and its complement
Why: successes over trials, then one minus.
\[ p' = \frac{300}{500} = 0.600, \quad q' = 0.400 \]
Check the condition
Why: both counts must exceed five.
\[ x = 300 > 5, \quad n - x = 200 > 5 \]
The z-score
Why: alpha 0.10, so 0.05 per tail and 0.95 to the left: invNorm(0.95, 0, 1).
\[ z_{0.05} = 1.645 \]
The product under the root
Why: multiply, then divide by n.
\[ \frac{(0.6)(0.4)}{500} = \frac{0.24}{500} = 0.00048 \]
The standard error
Why: the square root.
\[ \sqrt{0.00048} = 0.02191 \]
Keep at least four significant figures in the standard error. Rounding it to 0.02 here would give an error bound of 0.033 instead of 0.036, and the endpoints would move by three tenths of a point, enough to lose the mark.
The error bound
Why: z standard errors.
\[ \text{EBP} = (1.645)(0.02191) = 0.036 \]
The interval
Why: subtract and add.
\[ (0.600 - 0.036,\ 0.600 + 0.036) = (0.564,\ 0.636) \]
In context: we estimate with 90 percent confidence that the true percent of all students at the university who are registered voters is between 56.4 percent and 63.6 percent. The book's alternate wording says the same thing: between 56.4 and 63.6 percent of ALL students are registered voters.
Something surprising is hiding here. The level is lower than in Example 8.10, 90 against 95 percent, and the sample is the same size, yet the error bound is larger, 0.036 against 0.032. The next slide explains why.
Check against the calculator and the midpoint
Why: 1-PropZInt with x 300, n 500, level 0.90 prints (0.564, 0.636), and the centre must be p prime.
\[ \frac{0.564 + 0.636}{2} = 0.600 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 423 — Example 8.11
Picture it
The error bound depends on the data through one product, p prime times q prime. Plot that product across every possible proportion and the voter puzzle solves itself.
Figure (svg): A downward parabola of p prime times one minus p prime for p prime from 0 to 1, peaking at 0.25 when p prime is one half. Points are marked at 0.5 with 0.25, at 0.6 with 0.24, at 0.842 with 0.133 and at 0.9 with 0.09.
The curve is a parabola, highest in the middle. At 0.6 the product is 0.24, almost the maximum. At 0.842 it is only 0.133, because a lopsided split leaves less to be uncertain about. A survey where nearly everyone says yes has less variability than one that splits close to evenly.
\[ (0.6)(0.4) = 0.24 \quad > \quad (0.842)(0.158) = 0.133 \]
Take square roots and the voter survey's standard error is the root of 0.24 over 0.133, about 1.34 times the smartphone survey's, from the same n. That outweighs the smaller z.
\[ \sqrt{\frac{0.24}{0.133}} = 1.34 \quad > \quad \frac{1.96}{1.645} = 1.19 \]
Remember this shape: it comes back in Part 5, where its highest point, 0.25 at one half, becomes the safe assumption for planning a survey when you have no idea what the proportion will be.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 426 — the product p prime q prime is largest at one half
Estimation
National polls often interview about a thousand people about a question that splits close to evenly. Before computing, estimate the margin of error they can claim at 95 percent confidence.
Predict first
A poll of 1,000 people finds a proportion near one half. Roughly what is its 95 percent margin of error?
Correct: about 3 percentage points
Why: With a proportion of one half the product is 0.25, so the standard error is the square root of 0.25 over 1000, about 0.0158. Times 1.96 that is about 0.031, a margin of about three percentage points. That is why so many published polls quote plus or minus three points.
\[ 1.96\sqrt{\frac{(0.5)(0.5)}{1000}} = 1.96 \times 0.0158 = 0.031 \]
Remember the benchmark: about a thousand people, about 3 points. It lets you sanity-check any poll you read. A poll of 250 would have about twice that margin, since a quarter of the people doubles the error bound, and a poll of 4,000 about half of it.
Notice what does not appear: the size of the population. A poll of 1,000 is equally precise for a city of a million and a country of three hundred million, as long as the sample is random. Only n is in the formula.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 420 — polls quoted within three percentage points
Error analysis
Four students compute the 90 percent error bound for Example 8.11's voters. One is right. Find the fault in each of the others; every one of these mistakes appears on real papers.
Annotate
On: \( 3.60 \qquad 0.0008 \qquad 0.043 \qquad 0.036 \)
Two of the wrong answers fail a sense check. A margin above one is impossible for a proportion. A margin of 0.0008, under a tenth of a percentage point, would need millions of people; 500 cannot be that precise. Benchmark it: a thousand people give about 3 points, so 500 should give a little more than 3.
\[ 1.645\sqrt{\frac{0.24}{500}} = 1.645 \times 0.0219 = 0.036 \]
The wrong z is the dangerous one, because 0.043 looks perfectly reasonable. The only defence is to write the level and its z down before you start, and check them against each other at the end.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 423 — Example 8.11
Worked example
A student polls 600 students in a school district about new legislation on school uniforms; 480 are against it. Compute a 90 percent confidence interval for the true percent of students who are against the legislation, and interpret it.
\[ x = 480, \quad n = 600, \quad \text{CL} = 0.90 \]
The point estimate and complement
Why: successes over trials.
\[ p' = \frac{480}{600} = 0.80, \quad q' = 0.20 \]
Check the condition
Why: both counts exceed five.
\[ 480 > 5, \quad 120 > 5 \]
The z-score
Why: 90 percent, as in Example 8.11.
\[ z_{0.05} = 1.645 \]
The standard error
Why: product over n, then the root.
\[ \sqrt{\frac{(0.8)(0.2)}{600}} = \sqrt{\frac{0.16}{600}} = \sqrt{0.0002667} = 0.01633 \]
The error bound
Why: z standard errors.
\[ \text{EBP} = (1.645)(0.01633) = 0.0269 \]
The interval
Why: subtract and add.
\[ (0.800 - 0.027,\ 0.800 + 0.027) = (0.773,\ 0.827) \]
The margin is under 3 points, smaller than Example 8.11's 3.6, for two reasons that point the same way: a bigger sample, 600 against 500, and a proportion further from one half, so a smaller product, 0.16 against 0.24.
In context: we estimate with 90 percent confidence that between 77.3 percent and 82.7 percent of all students in the district are against the new uniform legislation.
Check the width
Why: the interval's length must be twice the error bound.
\[ 0.827 - 0.773 = 0.054 = 2(0.027) \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — Try It 8.11(a)
Worked example
In a sample of 300 students, 68 percent say they own an iPod and a smartphone. Compute a 97 percent confidence interval for the true percent of students who own both. This time the problem hands you a percent, not a count.
Convert the percent
Why: a proportion is a decimal.
\[ p' = 0.68, \quad q' = 0.32 \]
Recover the count, for the condition
Why: p prime times n.
\[ x = 0.68 \times 300 = 204, \quad n - x = 96 \]
Find alpha and the tail
Why: one minus the level, halved.
\[ \alpha = 1 - 0.97 = 0.03, \quad \frac{\alpha}{2} = 0.015 \]
The z-score
Why: 0.985 to the left: invNorm(0.985, 0, 1).
\[ z_{0.015} = 2.17 \]
Ninety-seven percent is not one of the four levels you know, so bracket it as a check: it lies between 95 and 98 percent, and 2.17 lies between their z-scores, 1.96 and 2.326.
The standard error
Why: product over n, then the root.
\[ \sqrt{\frac{(0.68)(0.32)}{300}} = \sqrt{\frac{0.2176}{300}} = \sqrt{0.0007253} = 0.02693 \]
The error bound
Why: z standard errors.
\[ \text{EBP} = (2.17)(0.02693) = 0.0584 \]
The interval
Why: subtract and add.
\[ (0.680 - 0.058,\ 0.680 + 0.058) = (0.622,\ 0.738) \]
In context: we estimate with 97 percent confidence that between 62.2 percent and 73.8 percent of all students own both an iPod and a smartphone.
Check the z by bracketing, and the range
Why: 97 percent must have a z between the 95 and 98 percent values, and both ends must lie in zero to one.
\[ 1.96 < 2.17 < 2.326, \quad 0 < 0.622 < 0.738 < 1 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — Try It 8.11(b)
Tweak it
The slider plots the 95 percent error bound, 1.96 times the square root of p prime q prime over n, against the sample proportion p prime from 0 to 1, for the sample size you choose.
Parameter explorer
Start at n equal to 100 and read the error bound at one half and at 0.9. Then set n to 400. What happens to the whole curve when n is multiplied by four, and where on the curve is the error bound always largest?
\[ \text{EBP} = 1.96\sqrt{p'q' / {n}} \]
Two lessons from the curve. Its peak never moves: whatever the sample size, a proportion of one half carries the largest error bound. And the whole curve scales together, so four times the sample halves the error bound everywhere.
\[ n \to 4n \quad \Rightarrow \quad \text{EBP} \to \frac{\text{EBP}}{2} \]
Near zero and one the curve dives to nothing. That looks like great precision, but it is the warning sign of Part 4: when p prime is near an end, the estimated standard error is unreliable, and a sample that happens to show no successes at all would claim an error bound of exactly zero.
Ranking
Five surveys from this section, each with its own count, size and level. Rank them from the smallest error bound to the largest. Think about all three ingredients, the level's z, the product p prime q prime, and n, before reaching for a calculator.
Put in order
Order the surveys from smallest error bound to largest.
Why: The error bounds are 0.027 for the uniforms, 0.032 for the smartphones, 0.036 for the voters, 0.058 for the 97 percent survey and 0.061 for the tablets. The two largest have the smallest samples; among the samples of 500 and 600, the proportion nearest one half has the widest margin.
\[ 0.027 < 0.032 < 0.036 < 0.058 < 0.061 \]
No single ingredient decides the order. The uniforms survey wins with the largest n and a fairly lopsided split. The voters lose to the smartphones despite the lower level, because 0.6 is so near one half. The tablets come last with the smallest sample and a product of 0.238.
The close pair is the 97 percent survey against the tablets. The 97 percent z, 2.17, is much bigger than 1.96, but the tablet sample is smaller and nearer one half, and those win by a little.
Section
Part 3
Concept
The interpretation sentence follows the same template as for a mean, with one change: the parameter is a percent or a proportion of a population, and the sentence should say so in the words of the problem.
| slot | what goes in it | Example 8.10 |
|---|---|---|
| the level | the confidence level as a percent | 95% confidence |
| the endpoints | both, usually converted to percents | between 81% and 87.4% |
| the population | who the proportion is of: all of them, not the sample | all adult residents of this city |
| the category | what counts as a success | have smartphones |
The book gives two wordings, and both are right. We estimate with 90 percent confidence that the true percent of all students that are registered voters is between 56.4 percent and 63.6 percent. Or: we estimate with 90 percent confidence that between 56.4 percent and 63.6 percent of all students are registered voters.
\[ (0.564,\ 0.636) \quad \longrightarrow \quad 56.4\% \text{ to } 63.6\% \]
The slot that goes wrong is the population. The sample's percentage is known exactly, 60 percent of the 500 surveyed; there is nothing to estimate about it. The interval is about all students at the university, and only random sampling earns you the right to talk about them.
A second sentence explains the level, about the method: ninety percent of all confidence intervals constructed in this way contain the true population percent.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 422-423 — the interpretations and explanations in Examples 8.10 and 8.11
Worked example
The smartphone interval was (0.810, 0.874) at 95 percent. Turn it into the two sentences the book writes, one slot at a time, so that nothing is lost on the way.
\[ (0.810,\ 0.874), \quad \text{CL} = 0.95, \quad n = 500 \]
Convert the endpoints to percents
Why: multiply each by one hundred.
\[ 0.810 \to 81.0\%, \qquad 0.874 \to 87.4\% \]
Name the population
Why: the survey drew randomly from the city's adults, so that is who the estimate covers.
\[ \text{all adult residents of this city} \]
Name the category
Why: what was counted as a success.
\[ \text{have smartphones} \]
Assemble the interpretation
Why: level, endpoints, population, category.
\[ \text{95\% confident: } 81\% \text{ to } 87.4\% \text{ of all adults} \]
Written out: we estimate with 95 percent confidence that between 81 percent and 87.4 percent of all adult residents of this city have smartphones.
Then the explanation of the level, which is about the procedure rather than this interval: ninety-five percent of the confidence intervals constructed in this way would contain the true value for the population proportion of all adult residents of this city who have smartphones.
Check the sentence against the numbers
Why: the midpoint of the percents must be the sample's percent, and the half-width the margin.
\[ \frac{81.0 + 87.4}{2} = 84.2\% = p', \quad \frac{87.4 - 81.0}{2} = 3.2 \text{ points} \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 422 — Example 8.10: interpretation and explanation of the 95 percent level
Trap
Example 8.11 gives (0.564, 0.636) at 90 percent. Two common sentences:
Wrong: between 56.4 and 63.6 percent of the 500 students surveyed are registered.
\[ 300/500 = 60\% \text{ exactly} \]
Wrong: there is a 90 percent chance that the true percent is between 56.4 and 63.6.
The sample's percent is known, 60 percent; nothing about it needs an interval. And the true percent is fixed, so this interval contains it or does not.
Right: we estimate with 90 percent confidence that between 56.4 and 63.6 percent of all students at the university are registered voters.
And the level: ninety percent of intervals built this way, from repeated random samples, contain the true percent.
The chance belongs to the method, before the sample is drawn; the estimate belongs to the whole population, not the 500.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 423 — Example 8.11: interpretation and alternate wording
Picture it
The book explains Example 8.11's level in one sentence: ninety percent of all confidence intervals constructed in this way contain the true population percent. Here is that sentence as an experiment, run in the deck's code. Suppose exactly 60 percent of all students are registered, draw 100 random samples of 500, and build the 90 percent interval from each.
Figure (svg): A hundred vertical bars side by side, each a 90 percent confidence interval for a proportion from a different simulated sample of 500 students drawn from a population in which exactly 0.6 are registered voters. A dashed yellow line marks 0.6. 94 bars, in blue, cross the line; 6, in red, lie wholly above or below it. The bars have almost the same length, about 0.072, because p prime barely changes the product near 0.6.
In this run 94 of the 100 intervals contain 0.6 and 6 miss. Unlike the intervals for a mean with sigma known, these do not all have exactly the same length: each one's error bound uses its own p prime. Near 0.6 the product hardly changes, so the lengths look equal to the eye.
\[ \text{about } 90 \text{ of } 100 \text{ intervals catch } p \]
The political science student holds exactly one of these bars and cannot see the yellow line. His bar, (0.564, 0.636), is either a hit or a miss. What he knows is that his method produces hits about nine times in ten, and that is all 90 percent confident means.
The count in one run is not exactly 90, for the same reason a hundred coin tosses rarely give exactly fifty heads. Over many runs the proportion of hits settles near 90 percent.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 423 — explanation of the 90 percent confidence level
Concept
A poll's margin of error sounds like a guarantee about everything that could go wrong with the poll. It is not. The error bound measures one thing only: how far a random sample's proportion typically lands from the population's, purely because of which people happened to be chosen.
\[ \text{EBP} = z\sqrt{\frac{p'q'}{n}} \quad \text{depends only on the level, } p' \text{ and } n \]
Look at the formula: nothing in it knows how the sample was gathered, how the question was worded, or who refused to answer. Those problems produce bias, a shift of the whole estimate in one direction, and a bigger sample does not shrink bias at all; it just measures the wrong thing more precisely.
| problem | what it does | covered by EBP? |
|---|---|---|
| random sampling variation | p′ lands above or below p by chance | yes |
| convenience or volunteer sample | the wrong population answers | no |
| people refuse or cannot be reached | the people who answer differ from those who do not | no |
| leading question wording | answers shift in one direction | no |
This is why every example in the section says randomly selected. Example 8.11's political science student surveyed 500 students; if he stood outside a voter-registration drive, his 60 percent would be biased upward, and the interval (0.564, 0.636) would be confidently wrong.
So read a margin of error as the smallest honest uncertainty the poll can have, the part that remains even if everything else was done perfectly.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 422-423 — randomly selected adults; a survey of 500 students
Discrimination
Six statements about Example 8.10's 95 percent interval for the proportion of the city's adults with smartphones. Sort each one, asking what it is describing: the population, the sample, individuals, or the method.
Sort into buckets
Sort each statement about the 95 percent interval (0.810, 0.874).
The statement about each adult is the new trap in this section. A proportion of the population is not a probability attached to each person. Every adult in the city either owns a smartphone or does not; the interval estimates what fraction do.
If a sentence would still be true without any sampling at all, as the one about the 500 surveyed would, it is not an interpretation of a confidence interval.
Worked example
An own example in the spirit of the homework. A newspaper reports that a candidate has 40 percent support with a margin of error of 3 percentage points at 95 percent confidence, but not how many voters were polled. Find p prime and the error bound, then estimate n.
\[ (0.37,\ 0.43), \quad \text{CL} = 0.95 \]
The point estimate
Why: the midpoint of the interval.
\[ p' = \frac{0.37 + 0.43}{2} = 0.40 \]
The error bound
Why: half the length.
\[ \text{EBP} = \frac{0.43 - 0.37}{2} = 0.03 \]
Write the error bound formula
Why: with q prime 0.60 and the 95 percent z.
\[ 0.03 = 1.96\sqrt{\frac{(0.4)(0.6)}{n}} \]
Isolate the root
Why: divide both sides by 1.96.
\[ \sqrt{\frac{0.24}{n}} = \frac{0.03}{1.96} = 0.015306 \]
Square both sides
Why: to remove the root.
\[ \frac{0.24}{n} = 0.00023428 \]
Solve for n
Why: divide 0.24 by that number.
\[ n = \frac{0.24}{0.00023428} = 1024.4 \]
So the poll interviewed about 1,025 voters. The published margin was rounded to a whole point, so you cannot recover n exactly; but about a thousand is clearly right, and it matches the benchmark from the estimation slide.
Many pollsters quote the margin using one half rather than their actual p prime, which would give n of about 1,067. Part 5 explains why that conservative choice is common.
Check by rebuilding the error bound
Why: 1,025 voters must give back a margin of 3 points.
\[ 1.96\sqrt{\frac{0.24}{1025}} = 1.96 \times 0.01530 = 0.0300 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 420 — a candidate with 40 percent of the vote within three percentage points
Real world
A hospital reviews 40 randomly selected surgical cases and finds 2 with a post-operative complication. A quality report quotes the rate as 5 percent and, using the ordinary formula, gives a 95 percent interval. Look at what the formula produces.
Discussion prompt
Compute the ordinary 95 percent interval for the complication rate from 2 of 40. What is wrong with its lower end, and which of the section's conditions did the report ignore?
Figure (svg): An axis from minus 0.05 to 0.20 with the part below zero shaded red and labelled impossible region. A red bar for the ordinary interval from 2 complications in 40 cases runs from minus 0.018 to 0.118, crossing into the red region. A blue bar for the plus-four interval, 4 of 44, runs from 0.006 to 0.176, entirely above zero.
The blue bar is the plus-four interval of Part 4, which stays inside the possible range. A negative endpoint is never a curiosity to round away quietly: it is the formula telling you its assumptions have failed.
In practice this matters. A hospital comparing itself with a benchmark rate of 10 percent would read the ordinary interval as clearly better; the adjusted one reaches 17.6 percent and says the data cannot tell.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — the interval can be used only if successes and failures both exceed five
Section
Part 4
Concept
The book admits that the interval for a proportion carries a certain amount of error. Because the true proportion is unknown, you are forced to estimate the standard deviation from p prime, and studies have shown that this estimate can be flawed, especially in small samples.
The fix is remarkably simple. Pretend you have four additional observations, two successes and two failures. Then run the ordinary procedure with the new counts.
\[ x \to x + 2, \qquad n \to n + 4 \]
\[ p'_{\text{plus-four}} = \frac{x + 2}{n + 4} \]
plus-four method — Add two successes and two failures to the observed data, then compute p′, q′, EBP and the interval exactly as usual with x + 2 and n + 4. Recommended when the confidence level is at least 90 percent and the sample size is at least ten.
Nothing else about the procedure changes: the same z, the same error bound formula, the same sentence. The book's reminder: you do not need to change the process, simply update the values of x and n to reflect the extra trials.
Computer studies have demonstrated the method's effectiveness, and the book gives the conditions plainly: use it when the confidence level desired is at least 90 percent and the sample size is at least ten.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 424-425 — the plus-four confidence interval for p
Picture it
A 95 percent interval promises to catch the true proportion 95 percent of the time. Is the promise kept? For 25 trials you can compute the true catch rate exactly from the binomial distribution, for every possible true proportion. The deck did that for both methods.
Figure (svg): Two jagged curves of the true coverage probability of a nominal 95 percent interval for a proportion from 25 trials, plotted against the true proportion from 0.02 to 0.98. The ordinary interval, in red, dips far below the dashed 0.95 line near both ends, falling to about 0.4 at the extremes and about 0.72 at 0.05 and 0.95, and wobbles below the line in many places in the middle. The plus-four interval, in blue, stays close to 0.95 across the whole range.
The red curve is the ordinary interval. Its catch rate is jagged and often well below 0.95: near a true proportion of 0.2 it is about 0.88, and near the ends it collapses below 0.75. The intervals are too narrow for their promise, because the estimated standard error is too small whenever p prime lands nearer the end than the truth.
\[ p = 0.20: \quad 0.884 \text{ (ordinary) } \quad 0.949 \text{ (plus-four)} \]
The blue curve is the plus-four interval. It stays close to the 0.95 line across nearly the whole range. Pulling the centre toward one half and adding four to n corrects the standard error in the direction it needs.
So the plus-four interval is not a fudge to make numbers look nicer. It is a better method, one whose 95 percent is closer to the truth, which is why the book recommends it for small samples.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — computer studies have demonstrated the effectiveness of this method
Step zero
Example 8.12 asks for a 95 percent interval for the proportion of statistics students who smoke, from 6 smokers in a random sample of 25, using the plus-four method. Before any arithmetic, take stock of what the problem has handed you and what you must change.
Discussion prompt
What is the parameter, in words? Is the plus-four method allowed here, and why is it a good idea? Which numbers will you actually put into the formula for x, n and z?
Taking stock heads off the two slips this example invites. The first is adjusting only x, or adjusting x and then using 25 in the standard error; both counts change, once, at the start. The second is reporting the interval as a statement about the 25 students, whose proportion is known exactly: 6 of 25, or 24 percent.
Notice also what the method does not change: the z-score comes from the level exactly as before, and the error bound formula is untouched. Only the inputs move.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — Example 8.12
Worked example
A random sample of 25 statistics students was asked whether they had smoked a cigarette in the past week; 6 said yes. Use the plus-four method to find a 95 percent confidence interval for the true proportion of statistics students who smoke.
\[ x = 6, \quad n = 25, \quad \text{CL} = 0.95 \]
Check the conditions for plus four
Why: level at least 90 percent, sample at least ten.
\[ 0.95 \ge 0.90, \quad 25 \ge 10 \]
Add two successes and two failures
Why: the plus-four counts.
\[ x = 6 + 2 = 8, \quad n = 25 + 4 = 29 \]
The adjusted point estimate
Why: the new successes over the new trials.
\[ p' = \frac{8}{29} = 0.276, \quad q' = 0.724 \]
The z-score
Why: 95 percent.
\[ z_{0.025} = 1.96 \]
The standard error
Why: with the adjusted counts throughout, including n of 29.
\[ \sqrt{\frac{(0.276)(0.724)}{29}} = \sqrt{0.006888} = 0.0830 \]
The error bound
Why: z standard errors.
\[ \text{EBP} = (1.96)(0.0830) = 0.163 \]
The interval
Why: subtract and add.
\[ (0.276 - 0.163,\ 0.276 + 0.163) = (0.113,\ 0.439) \]
In context: we are 95 percent confident that the true proportion of all statistics students who smoke cigarettes is between 0.113 and 0.439, that is between 11.3 and 43.9 percent.
The interval is wide because 25 students is a small sample. That honesty is the point: the ordinary interval from the same data would claim a starting point of 7.3 percent, which the coverage picture says is too optimistic.
Check against the calculator
Why: 1-PropZInt with x 8, n 29 and level 0.95 prints (0.113, 0.439); the midpoint must be the adjusted p prime.
\[ \frac{0.113 + 0.439}{2} = 0.276 = \frac{8}{29} \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 424-425 — Example 8.12
Picture it
Put the two intervals for Example 8.12's smokers on one line: the ordinary one from 6 of 25, and the plus-four one from 8 of 29.
Figure (svg): Above a proportion axis from 0 to 0.5, a red bar for the ordinary interval centred at 0.240 from 0.073 to 0.407, and below it a blue bar for the plus-four interval centred at 0.276 from 0.113 to 0.439. A yellow arrow runs from the red centre to the blue centre, toward a dashed line at one half.
\[ \text{ordinary: } \frac{6}{25} = 0.240 \pm 0.167 \qquad \text{plus-four: } \frac{8}{29} = 0.276 \pm 0.163 \]
Two things moved. The centre shifted from 0.240 up to 0.276, toward one half, because adding equal numbers of successes and failures always drags a proportion toward the middle. And the error bound barely changed, 0.167 against 0.163, since the product rose a little while n rose too.
The shift matters most at the low end. The ordinary interval starts at 0.073; the plus-four interval starts at 0.113. The ordinary interval is the one that more often misses, and it misses by being too close to zero.
On a large sample the same four observations would change almost nothing. Add two and two to 1,000 people and the proportion moves by a fraction of a point. That is why the method is aimed at small samples, and why using it on large ones does no harm.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — Example 8.12
Prediction
A random sample of 30 shoppers finds that 3 used a coupon, so the ordinary p prime is 0.10. Before computing, decide where the plus-four estimate will land.
Predict first
With 3 coupon users among 30 shoppers, how does the plus-four estimate compare with the ordinary one of 0.10?
Correct: it is larger, moved toward one half
Why: Five successes in 34 trials is about 0.147, up from 0.10. Adding two successes and two failures adds exactly half of the new observations as successes, so the estimate is pulled toward one half. It can never cross one half: a proportion below one half stays below it.
\[ \frac{3 + 2}{30 + 4} = \frac{5}{34} = 0.147 \quad > \quad 0.10 \]
A proportion of exactly one half does not move at all: 15 of 30 becomes 17 of 34, still one half. A proportion above one half moves down toward it. The pull is always toward the middle, and stronger for small samples and for proportions far from one half.
That is precisely the direction the coverage picture asked for. The ordinary interval misses most often near the ends, and the adjustment moves the centre away from the ends.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — two of these observations are successes and two are failures
Worked example
Of a random sample of 65 first-year students at State University, 31 have declared a major. Use the plus-four method to find a 96 percent confidence interval for the true proportion of first-year students who have declared a major.
\[ x = 31, \quad n = 65, \quad \text{CL} = 0.96 \]
The plus-four counts
Why: two successes and two failures.
\[ x = 31 + 2 = 33, \quad n = 65 + 4 = 69 \]
The adjusted point estimate
Why: the new counts.
\[ p' = \frac{33}{69} = 0.478, \quad q' = 0.522 \]
Find alpha and the tail
Why: a 96 percent level.
\[ \alpha = 0.04, \quad \frac{\alpha}{2} = 0.02 \]
The z-score
Why: 0.98 to the left: invNorm(0.98, 0, 1).
\[ z_{0.02} = 2.054 \]
A 96 percent level is not on the familiar list, so the z must be computed rather than remembered. The tail is 0.02, so you give the calculator 0.98, the area to the left of the upper cut. A table would show 0.9798 at 2.05 and 0.9803 at 2.06.
The standard error
Why: product over 69, then the root.
\[ \sqrt{\frac{(0.478)(0.522)}{69}} = \sqrt{0.003616} = 0.0601 \]
The error bound
Why: z standard errors.
\[ \text{EBP} = (2.054)(0.0601) = 0.124 \]
The interval
Why: subtract and add.
\[ (0.478 - 0.124,\ 0.478 + 0.124) = (0.355,\ 0.602) \]
Here the adjustment barely moved the centre: 31 of 65 is 0.477 and 33 of 69 is 0.478, because the proportion was already near one half. The plus-four method matters little in the middle and most near the ends.
In context: we are 96 percent confident that between 35.5 percent and 60.2 percent of all first-year students at State University have declared a major.
Check the z by bracketing and the midpoint
Why: 96 percent sits between 95 and 98 percent, and the centre must be 33 over 69.
\[ 1.96 < 2.054 < 2.326, \quad \frac{0.355 + 0.602}{2} = 0.478 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 425 — Try It 8.12
Worked example
A marketing research firm asked 50 adults aged 18-29 about buying an electric vehicle; 13 said they would consider it. Use the plus-four method to find a 90 percent confidence interval for the true proportion of adults aged 18-29 who would consider the purchase.
\[ x = 13, \quad n = 50, \quad \text{CL} = 0.90 \]
The plus-four counts
Why: two successes and two failures.
\[ x = 13 + 2 = 15, \quad n = 50 + 4 = 54 \]
The adjusted point estimate
Why: the new counts.
\[ p' = \frac{15}{54} = 0.278, \quad q' = 0.722 \]
Both conditions for plus four hold: the level is 90 percent and the sample is 50. Without the adjustment the condition would also pass, with 13 successes and 37 failures, but the sample is small enough that the adjustment is still worth making.
The z-score
Why: alpha 0.10, so 0.05 per tail.
\[ z_{0.05} = 1.645 \]
The standard error
Why: product over 54, then the root.
\[ \sqrt{\frac{(0.278)(0.722)}{54}} = \sqrt{0.003715} = 0.0610 \]
The error bound
Why: z standard errors.
\[ \text{EBP} = (1.645)(0.0610) = 0.100 \]
The interval
Why: subtract and add.
\[ (0.278 - 0.100,\ 0.278 + 0.100) = (0.178,\ 0.378) \]
In context: we are 90 percent confident that between 17.8 percent and 37.8 percent of adults aged 18-29 would consider the purchase of an electric vehicle.
Compare the ordinary interval from 13 of 50: centre 0.26, error bound 0.102, interval (0.158, 0.362). The plus-four interval sits about two points higher, pulled toward one half, with almost the same width.
Check the width and the calculator
Why: the length must be twice the error bound, and 1-PropZInt with x 15, n 54 and level 0.90 prints (0.178, 0.378).
\[ 0.378 - 0.178 = 0.200 = 2(0.100) \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 425-426 — Example 8.13
Worked example
The electric-vehicle survey also interviewed people by phone. In the end 588 adults aged 18-29 answered, and 159 said they would consider buying one. Use the plus-four method for a 90 percent interval and compare it with Example 8.13.
\[ x = 159, \quad n = 588, \quad \text{CL} = 0.90 \]
The plus-four counts
Why: two successes and two failures.
\[ x = 161, \quad n = 592 \]
The adjusted point estimate
Why: the new counts.
\[ p' = \frac{161}{592} = 0.272, \quad q' = 0.728 \]
The book asks for plus four even here, so follow it, but notice how little it changes. Adding two and two to nearly six hundred people moves the estimate from 0.2704 to 0.2720, less than a fifth of a point.
The standard error
Why: product over 592, then the root.
\[ \sqrt{\frac{(0.272)(0.728)}{592}} = \sqrt{0.0003345} = 0.01829 \]
The error bound
Why: the 90 percent z.
\[ \text{EBP} = (1.645)(0.01829) = 0.030 \]
The interval
Why: subtract and add.
\[ (0.272 - 0.030,\ 0.272 + 0.030) = (0.242,\ 0.302) \]
Figure (svg): Two interval bars above an axis from 0.10 to 0.45. Orange, from Example 8.13 with 54 adjusted trials: centred at 0.278, from 0.178 to 0.378. Blue, from Try It 8.13 with 592 adjusted trials: centred at 0.272, from 0.242 to 0.302, lying inside the orange bar.
The book asks for a comparison. The centres nearly agree, 0.278 and 0.272, so the larger sample confirms the smaller one. The new interval is about a third as wide, 0.060 against 0.200, and lies entirely inside the old one. With 588 people the plus-four adjustment hardly matters: the ordinary p prime, 159 over 588, is 0.270.
In context: we are 90 percent confident that between 24.2 percent and 30.2 percent of adults aged 18-29 would consider buying an electric vehicle.
Check the square-root law
Why: about eleven times the adjusted sample should shrink the margin by about the root of eleven.
\[ \frac{0.100}{0.030} = 3.3 \approx \sqrt{\frac{592}{54}} = 3.3 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 426 — Try It 8.13
Estimation
Two surveys each find 15 percent in the category of interest: one asked 40 people, the other 1,000. Both apply the plus-four adjustment. Estimate how far it moves each estimate before computing.
Predict first
Six of 40 and 150 of 1000 both give 0.15. About how far does the plus-four adjustment move each estimate?
Correct: about 3 points for the 40, and well under a point for the 1000
Why: Eight of 44 is 0.182, a move of more than three points from 0.15, while 152 of 1004 is 0.1514, a move of about a seventh of a point. Four observations are a tenth of a sample of 40 but only a quarter of a percent of a sample of 1000.
\[ \frac{6 + 2}{40 + 4} = 0.182, \qquad \frac{150 + 2}{1000 + 4} = 0.1514 \]
The adjustment fades as n grows, which is exactly right: large samples do not need rescuing, and the ordinary interval is already reliable for them. Using plus four on a big survey costs nothing, and using it on a small one fixes a real problem.
For the large survey the two 95 percent intervals agree to three decimals in width, about 0.022 either side, and their centres differ by about a seventh of a point. You could report either.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — the new sample size is n + 4
Trap
Example 8.12 again, 6 of 25. A first attempt adds two to x and leaves n alone:
\[ p' = \frac{6 + 2}{25} = 0.32 \]
Wrong. That adds two successes and no failures: it is not plus four, it is plus two successes.
\[ (0.137,\ 0.503) \quad \times \]
Four observations are added: two successes and two failures. Both counts change.
\[ p' = \frac{6 + 2}{25 + 4} = \frac{8}{29} \]
\[ = 0.276 \]
The wrong version drags the estimate up by eight points rather than nudging it toward one half. The same slip in the error bound, using n of 25 instead of 29, widens the interval too. Change x and n once, at the start, and use the new pair everywhere after.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 424 — the new sample size is n + 4, and the new count of successes is x + 2
Counterexample
The ordinary interval can fail in a stranger way than running past zero. Suppose a random sample of 30 new phones is tested and none has a defective battery. Try to build the 95 percent interval for the defect rate.
Discussion prompt
With 0 defects among 30 phones, compute the ordinary 95 percent interval. Then compute the plus-four interval. Which one gives a believable answer, and why?
This is the clearest case for the adjustment. Seeing no successes does not mean the proportion is zero; it means the proportion is small. Any method that turns no evidence into perfect certainty is broken.
The book's condition would have caught this before any arithmetic: zero successes is not more than five, so the ordinary interval was never allowed.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 421, 424 — the success condition, and the plus-four interval
Section
Part 5
Intuition
Before a survey, researchers decide how large a margin of error they can accept and how confident they want to be. The error bound formula links those to n, so solve it for n, one move at a time.
\[ \text{EBP} = z\sqrt{\frac{p'q'}{n}} \]
Divide both sides by z, so the root stands alone.
\[ \frac{\text{EBP}}{z} = \sqrt{\frac{p'q'}{n}} \]
Square both sides to remove the root.
\[ \frac{\text{EBP}^{2}}{z^{2}} = \frac{p'q'}{n} \]
Flip both sides and multiply by p prime q prime to bring n to the top.
\[ n = \frac{z^{2}\,p'q'}{\text{EBP}^{2}} \]
This is the book's sample-size formula. It has the same structure as the one for a mean in Section 8.1, with p prime q prime in the place of sigma squared. And it has the same unwelcome feature: it needs p prime, the result of a survey that has not yet been run.
Read off what drives the sample size. It grows with the square of z, so higher confidence costs more. It grows with p prime times q prime, so evenly split populations cost most. And it grows with one over the square of the error bound, so halving the margin multiplies the sample by four.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 426 — calculating the sample size n
Concept
To find n you need p prime, but you have not collected any data yet. The book's way out: since p prime and q prime are multiplied together, choose them both to be 0.5, because that makes the product as large as it can possibly be.
| p′ | q′ | p′q′ |
|---|---|---|
| 0.5 | 0.5 | 0.25 (largest) |
| 0.6 | 0.4 | 0.24 |
| 0.7 | 0.3 | 0.21 |
| 0.8 | 0.2 | 0.16 |
| 0.9 | 0.1 | 0.09 |
The book checks the claim by trial, as in this table: every other split gives a smaller product. The largest product gives the largest n, so a survey planned with 0.5 is large enough whatever the true proportion turns out to be.
\[ n = \frac{z^{2}(0.5)(0.5)}{\text{EBP}^{2}} = \frac{0.25\,z^{2}}{\text{EBP}^{2}} \]
The cost of being safe is small near the middle. Any true proportion between 0.3 and 0.7 has a product of at least 0.21, within 16 percent of the maximum, so the conservative plan is rarely much bigger than it needed to be.
If an earlier study gives a reliable estimate far from one half, you may use it and survey fewer people. But that is a bet, and the trap later in this part shows what it costs when it loses.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 426 — Example 8.14: we make them both equal to 0.5
Matching
Sample-size problems are written in plain words, not symbols. Match each phrase from Example 8.14 and its relatives to the quantity it sets in the formula.
Match the pairs
Why: Within three points is the error bound, 0.03. Ninety percent confidence gives the z of 1.645. An unknown proportion means using one half for both. The counts of a finished survey give x and n. And the question how many is asking for n, which is always rounded up.
Percentage points become a decimal error bound: three points is 0.03, not 3. Putting 3 into the formula gives a sample size of less than one person, which is a quick sign of the slip.
The counts of a finished survey do not belong in a planning problem at all; they belong to the interval you build afterwards. If a planning problem does give you an earlier p prime, it replaces the one half, and you should ask how far you trust it.
\[ \text{3 percentage points} = 0.03 \]
The other phrase to watch is within. A margin of 3 points means 3 points on each side; if a problem instead fixes the total width of the interval, halve it before using the formula.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 426 — Example 8.14
Worked example
A mobile phone company wants the current percentage of its customers aged 50 and over who use text messaging. How many should it survey to be 90 percent confident that the sample proportion is within three percentage points of the true proportion?
\[ \text{EBP} = 0.03, \quad \text{CL} = 0.90, \quad p' \text{ unknown} \]
The z-score
Why: 90 percent.
\[ z_{0.05} = 1.645 \]
Choose p prime and q prime
Why: unknown, so use the largest product.
\[ p' = q' = 0.5, \quad p'q' = 0.25 \]
The company has no survey yet, so it has no p prime. Using one half is not a claim that half the customers text; it is a promise that the sample will be big enough for any proportion, because no proportion has a larger product than one half.
Write the formula
Why: with the numbers in place.
\[ n = \frac{z^{2}\,p'q'}{\text{EBP}^{2}} = \frac{(1.645)^{2}(0.25)}{(0.03)^{2}} \]
Square each piece
Why: 1.645 squared and 0.03 squared.
\[ n = \frac{(2.706025)(0.25)}{0.0009} \]
Multiply the numerator
Why: times one quarter.
\[ n = \frac{0.67650625}{0.0009} \]
Divide
Why: the book's value.
\[ n = 751.67 \]
With the exact z, 1.6449, the quotient is 751.54. Either way the answer is not a whole number, and a survey must reach a whole number of people, so round up to 752.
Figure (svg): A curve of the required sample size, 1.645 squared times 0.25 over the target error bound squared, for error bounds from 0.015 to 0.08. It falls from about 3007 at 0.015 to about 106 at 0.08. Points mark 0.03 with 751.7, rounded up to 752, and 0.05 with 270.6, rounded up to 271.
In context: the company should survey 752 customers aged 50 and over to be 90 percent confident that the sample proportion is within three percentage points of the true proportion of all such customers who use text messaging.
Check by putting n = 752 back in
Why: the error bound, with p prime 0.5, must be at most 0.03.
\[ 1.645\sqrt{\frac{0.25}{752}} = 1.645 \times 0.018233 = 0.02999 \le 0.03 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 426-427 — Example 8.14
Trap
Example 8.14 gives n equal to 751.67. A common habit is to drop the decimals, since you cannot survey part of a person:
\[ n = 751.67 \to 751 \]
Wrong. With 751 people the error bound is:
\[ 1.645\sqrt{\frac{0.25}{751}} = 0.030013 \]
That is just over the promised 3 points.
The formula gives the smallest n that exactly meets the target, so any whole number below it falls short. Always round up.
Figure (svg): Five points of the error bound 1.645 times the square root of 0.25 over n for n from 749 to 753, falling from 0.03005 to 0.02997, with a dashed red line at the target 0.03. Points for 749, 750 and 751 are red and above the line; 752 and 753 are green and below it.
\[ n = 752: \]
\[ \text{EBP} = 0.029993 \le 0.03 \]
Ordinary rounding happens to give 752 here, but with a decimal below one half, such as the 216.09 of Section 8.1, it would round down too. The book says it directly: round the answer to the next higher value. Rounding up costs at most one extra interview; rounding down breaks the promise the study was designed to keep.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 427 — round the answer to the next higher value
Worked example
An internet marketing company wants the current percentage of customers who click on ads on their smartphones. How many should it survey to be 90 percent confident that the estimated proportion is within five percentage points of the true proportion?
\[ \text{EBP} = 0.05, \quad z = 1.645, \quad p' = q' = 0.5 \]
Substitute
Why: into the sample size formula.
\[ n = \frac{(1.645)^{2}(0.25)}{(0.05)^{2}} \]
Square each piece
Why: the same numerator as Example 8.14.
\[ n = \frac{0.67650625}{0.0025} \]
Divide
Why: a larger margin, a smaller denominator's square.
\[ n = 270.6 \]
Round up
Why: to the next whole number.
\[ n = 271 \]
The decimal is 0.6, so ordinary rounding and rounding up agree this time. They will not always agree, so make rounding up a fixed habit rather than a judgment made fresh in each problem.
In context: the company should survey 271 customers to be 90 percent confident that its sample proportion is within five percentage points of the true proportion who click on ads.
Compare with Example 8.14: loosening the margin from 3 points to 5 cut the survey from 752 to 271. The ratio is five thirds squared, about 2.78, because the margin is squared in the denominator.
\[ \frac{751.67}{270.6} = 2.78 = \left(\frac{5}{3}\right)^{2} \]
Check both neighbours
Why: 271 must meet the target and 270 must not.
\[ 1.645\sqrt{\frac{0.25}{271}} = 0.04996 \le 0.05, \quad 1.645\sqrt{\frac{0.25}{270}} = 0.05006 > 0.05 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 427 — Try It 8.14
Trap
A planner for Example 8.14 guesses that about 30 percent of older customers text, and uses it:
\[ n = \frac{(1.645)^{2}(0.3)(0.7)}{(0.03)^{2}} \]
\[ = 631.4 \Rightarrow 632 \]
Risky. If the true proportion is 0.5, the finished survey's margin is:
\[ 1.645\sqrt{\frac{0.25}{632}} = 0.0327 \]
The guess saved 120 interviews and broke the promise of 3 points.
Figure (svg): A curve of the required sample size for a 3 point margin at 90 percent confidence, 3006.7 times p times one minus p, against the assumed proportion from 0 to 1. It peaks at 751.7 when p is one half. A point at 0.3 gives 631.4. A dashed line at 752 marks the conservative plan.
Any assumed proportion other than one half gives a smaller n, so it only works if the guess is right or further from one half than the truth. Without solid prior evidence, plan with 0.5: n of 752 meets the 3 point margin whatever the proportion turns out to be.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 426 — the largest possible product gives the largest n
Worked example
An own example showing when a guess other than one half is legitimate. A health department knows from five years of large surveys that about 10 percent of adults in its region smoke, never above 12 percent. It wants a new estimate within 2 percentage points at 95 percent confidence. How many adults should it survey?
\[ \text{EBP} = 0.02, \quad z = 1.96, \quad \text{prior estimate } 0.10, \text{ never above } 0.12 \]
The conservative plan, for comparison
Why: with one half.
\[ n = \frac{(1.96)^{2}(0.25)}{(0.02)^{2}} = \frac{0.9604}{0.0004} = 2401 \]
That is a large and expensive survey, and the department has good evidence that the proportion is nowhere near one half. Using that evidence is fair, but use the value nearest one half that the evidence allows, 0.12, not the typical 0.10, so the plan still covers the worst plausible case.
The product at the worst plausible proportion
Why: 0.12 is the value nearest one half.
\[ (0.12)(0.88) = 0.1056 \]
Substitute
Why: into the formula.
\[ n = \frac{(1.96)^{2}(0.1056)}{(0.02)^{2}} = \frac{(3.8416)(0.1056)}{0.0004} \]
Multiply and divide
Why: then round up.
\[ n = \frac{0.40567}{0.0004} = 1014.2 \Rightarrow 1015 \]
Planning with 0.10 itself would have given 864.4, so 865 people; that plan would miss the 2 point target if the proportion had drifted up to 12 percent, giving a margin of about 2.2 points.
In context: surveying 1,015 adults gives 95 percent confidence of being within 2 points of the true smoking rate, as long as that rate is no more than 12 percent, at well under half the cost of the conservative plan.
Check by putting n = 1015 back in
Why: with the worst plausible proportion, the margin must be at most 0.02.
\[ 1.96\sqrt{\frac{0.1056}{1015}} = 1.96 \times 0.010200 = 0.01999 \le 0.02 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 426 — the sample size formula with an estimated proportion
Edge cases
Example 8.14 needed 752 customers for a 3 point margin at 90 percent. The company's managers start asking for more. Before computing, guess how much bigger each request makes the survey.
Discussion prompt
Using p prime of 0.5, find the sample size for (a) a 1.5 point margin at 90 percent, (b) a 3 point margin at 95 percent, and (c) a 2 point margin at 95 percent. Which request is most expensive, and why?
Remember the asymmetry. The margin is squared in the denominator, so every halving quadruples the cost. The level enters only through z, which grows slowly. That is why national polls settle for 3 points at 95 percent with about a thousand people: one point would need nine times as many.
\[ n_{1\text{ pt}} = 9 \times n_{3\text{ pt}} \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 426-427 — the sample size formula
Comparison
Section 8.1 and this section build intervals of exactly the same shape. Fill in the blanks to collect what differs. Type the plain words or symbols shown in the other rows.
Comparison matrix
| piece | population mean (σ known) | population proportion |
|---|---|---|
| point estimate | x̄ | p' |
| multiplier | z for the level | z |
| standard error | σ/√n | √(p′q′/n) |
| spread needs to be given? | yes, σ | no |
| planning value when unknown | σ from a pilot study | 0.5 |
The z column is the reassuring one: nothing about turning a level into a multiplier has changed. Whatever you learned about 1.645, 1.96 and 2.576 carries straight over.
The spread row is the structural difference worth carrying out of the chapter. For a mean, the centre and the spread are separate numbers, and the spread has to come from somewhere. For a proportion, the centre fixes the spread, so one number does both jobs; that is also why the planning value is one half, the proportion with the largest spread.
\[ \text{mean: } n = \frac{z^{2}\sigma^{2}}{\text{EBM}^{2}} \qquad \text{proportion: } n = \frac{z^{2}p'q'}{\text{EBP}^{2}} \]
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, p. 421 — similar to the error bound formula for a mean
Section
Part 6
Pattern
Every interval in this section came from the same six moves. The plus-four adjustment, when you use it, happens before the second step and changes nothing after it.
Figure (svg): Six boxes in a row joined by arrows: recognise yes or no data; compute p prime and q prime; check that the successes and failures both exceed five; find z with area one minus alpha over two to its left; compute the error bound z times the root of p prime q prime over n; write the sentence as a percent of the population, highlighted in yellow. Under the conditions box a note reads small n, plus-four first.
\[ p' \pm z_{\alpha/2}\sqrt{\frac{p'q'}{n}}, \qquad n = \frac{z^{2}(0.25)}{\text{EBP}^{2}} \text{ (round up)} \]
Three quick checks catch most errors: both endpoints between zero and one, the midpoint equal to your p prime, and a margin of about 3 points for every thousand or so people near one half.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 421-427 — the whole procedure
Check
A fresh survey. Decide p prime, the z and the standard error in that order before choosing.
Check your understanding
In a random sample of 400 commuters, 126 ride a bicycle at least once a week. Which is the 90 percent confidence interval for the proportion of all commuters who do?
Answer: A
Why: The sample proportion is 126 over 400, which is 0.315, and q prime is 0.685. Their product over 400 is 0.000539, whose square root is 0.0232. The 90 percent z is 1.645, so the error bound is 0.038, and the interval runs from 0.277 to 0.353.
\[ p' = 0.315, \quad \sqrt{\frac{(0.315)(0.685)}{400}} = 0.0232, \quad \text{EBP} = 1.645(0.0232) = 0.038 \]
Check it the usual ways: the midpoint of 0.277 and 0.353 is 0.315, both ends are between zero and one, and a margin near 4 points from 400 people fits the benchmark of 3 points per thousand.
The condition holds comfortably: 126 cyclists and 274 others. Every wrong option is one of this lesson's slips: the wrong level's z, a missing square root, and percentages typed into the formula. Two of them are caught by a range check alone.
In context: we estimate with 90 percent confidence that between 27.7 and 35.3 percent of all commuters ride a bicycle at least once a week.
Check
This check is about the adjustment itself. Change the counts first, then run the ordinary procedure.
Check your understanding
A random sample of 100 adults finds 10 who have tried a meal-kit service. Which is the 95 percent plus-four interval for the proportion of all adults who have?
Answer: A
Why: Add two successes and two failures: 12 out of 104, which is 0.1154. The standard error is the square root of 0.1154 times 0.8846 over 104, about 0.0313, and 1.96 times that is 0.061. The interval is 0.054 to 0.177.
\[ p' = \frac{12}{104} = 0.1154, \quad \text{EBP} = 1.96\sqrt{\frac{(0.1154)(0.8846)}{104}} = 0.061 \]
The adjusted centre, 0.115, sits a little above 0.10, nudged toward one half as it always is. The width is almost the same as the ordinary interval's; what changed is mainly where it sits.
Write the adjusted counts down first and cross out the originals, so that the old n of 100 cannot slip back into the standard error. That one habit prevents the two most common wrong answers here.
The conditions hold: the level is at least 90 percent and the sample at least ten. With only 10 successes the ordinary interval is legal too, but the plus-four one keeps its 95 percent promise more reliably.
Check
Use the sample size formula with the conservative proportion, then decide how to round.
Check your understanding
A student government wants to estimate the proportion of students who support a new fee, to within 4 percentage points with 95 percent confidence. Nothing is known about the proportion. How many students should it survey?
Answer: A
Why: With z of 1.96 and one half for both proportions, n is 1.96 squared times 0.25 over 0.04 squared, which is 0.9604 over 0.0016, or 600.25. Rounding up gives 601, the smallest whole number that meets the 4 point margin.
\[ n = \frac{(1.96)^{2}(0.25)}{(0.04)^{2}} = \frac{0.9604}{0.0016} = 600.25 \Rightarrow 601 \]
Verify by putting 601 back in: 1.96 times the square root of 0.25 over 601 is 0.03998, just inside 0.04. With 600 it is 0.04001, just outside, which is the whole case for rounding up.
Notice that the answer does not depend on how many students attend the school. As with the margin of a poll, only the level, the margin and the planning proportion enter the formula; a school of two thousand and one of twenty thousand need the same sample.
In context: surveying 601 randomly chosen students gives 95 percent confidence that the sample proportion is within 4 points of the true proportion who support the fee, whatever that proportion is.
Explain it to yourself
The idea that makes this section different from 8.1 and 8.2 is short to state and easy to say wrongly. Try it in plain words before reading the model answer.
Discussion prompt
A classmate asks: for a mean we needed sigma or s to get the error bound, so where is the standard deviation in the proportion formula, and why do people plan surveys with one half? Explain both in a few sentences.
A strong answer names the link, the centre of a yes or no variable fixes its spread, and then uses that same link for planning. If your explanation said the standard deviation is assumed or ignored, it has missed the point: it is not ignored, it is determined.
If you can also say why proportions near one half are hardest to estimate, you have the whole picture: the most evenly split population has the most variable answers.
Exit ticket
A random sample of 200 voters finds that 90 support a ballot measure. The campaign wants to claim majority support.
Discussion prompt
Find (a) the 95 percent confidence interval for the proportion of all voters who support the measure, (b) whether it supports a claim of majority support, and (c) how many voters a follow-up poll needs for a 3 point margin at 95 percent, with nothing assumed about the proportion.
The marks are in three places: working in decimals, reading the interval as a statement about all voters rather than about 0.5 itself, and rounding the planned sample up.
A quick self-check: 200 voters near one half should give a margin around 7 points, since a fifth of a thousand more than doubles the 3 point benchmark. The interval is 6.9 points either side.
Recap
A proportion problem has yes or no data and no mention of a mean. The point estimate is p prime, the number of successes over the sample size; q prime is one minus it.
| piece | formula | Example 8.10 |
|---|---|---|
| point estimate | p′ = x/n, q′ = 1 − p′ | 0.842, 0.158 |
| condition | x > 5 and n − x > 5 | 421 and 79 |
| z | invNorm(1 − α/2) | 1.96 |
| error bound | EBP = z·√(p′q′/n) | 0.032 |
| interval | (p′ − EBP, p′ + EBP) | (0.810, 0.874) |
| plus four | x + 2, n + 4; CL ≥ 90%, n ≥ 10 | smokers: 8/29 |
| sample size | n = z²p′q′/EBP², p′ = 0.5, round up | Example 8.14: 752 |
No population standard deviation appears, because a yes or no variable's spread is fixed by its proportion. That spread is largest at one half, which makes proportions near one half the hardest to estimate and one half the safe planning value.
Every interval ends in a sentence about a percent of the whole population, in the words of the problem; the confidence level describes the method. Introductory Business Statistics builds the same interval with business data if you want more practice.
OpenStax Introductory Statistics 2e, §8.3 A Population Proportion §8.3, pp. 420-427 — the whole section
OpenStax Introductory Business Statistics 2e, §8.3 A Confidence Interval for A Population Proportion §8.3 A Confidence Interval for A Population Proportion — the same interval in business contexts
Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.