The binomial's independence characteristic relaxed. A hypergeometric experiment samples without replacement from two groups, so each pick changes what remains, the picks are not independent, and there is no single success probability to raise to a power. The count of items drawn from the group of interest has a hypergeometric distribution, written X follows H of r, b and n, and its probabilities are computed by counting combinations: the ways of choosing x from the group of interest, times the ways of choosing the rest from the second group, over the ways of choosing the sample at all. The possible values of X are limited by the group sizes as well as by the sample size, and the mean is the sample size times the proportion in the group of interest.
Subject: Statistics · 66 slides · symbolic lesson
Open the interactive version of this deck
Title
Statistics · Chapter 4 — Discrete Random Variables
Counting what lands in a sample when nothing is put back
Objectives
The binomial distribution of Section 4.3 needed independent trials with the same probability of success every time. Deal cards, pick a committee or pull laptops off a pallet for inspection, and that condition fails: nothing goes back, so each pick changes what is left. This section handles exactly that case.
One question runs underneath all of it: what happens to the pool after each pick? If the answer is nothing, because the item goes back or the pool is effectively endless, you are in binomial territory. If the pool shrinks and its make-up shifts, you are here.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, pp. 247-250 — the section this deck covers
Warm-up
The book's own Example 4.25 committee: four people are chosen at random from six men and five women. Use the multiplication rule from Chapter 3, one pick at a time.
Discussion prompt
What is the probability that the first two people chosen are both men? Compare it with the answer you would get if each chosen person were put back before the next pick.
Ask the question once more for the third pick, given that the first two were men. Four men remain among nine people, so the chance is four ninths, about 0.44, well down from the six elevenths, about 0.55, you started with. Each man chosen makes the next man less likely.
Keep both lines in view. The first is what really happens when a committee is chosen; the second is what a binomial model would pretend. The gap between 0.2727 and 0.2975 is small after two picks, and the whole of this section is about what to do when that gap matters.
Section
Part 1
Concept
The book's opening example: a softball team of ten players is chosen from a combined group of 11 men and 13 women. Look at the chances for the first two picks.
Figure (svg): Three panels, each a grid of 24 circles: 11 blue for men and 13 orange for women. The first panel is the full pool. In the second one orange circle is dashed, picked, leaving 11 men and 12 women; in the third one blue circle is dashed, leaving 10 men and 13 women.
Before anyone is chosen, a woman comes first with probability thirteen out of twenty-four. After that, only 23 people are left, and who they are depends on the first pick.
\[ P(\text{man second} \mid \text{woman first}) = \frac{11}{23} \approx 0.478 \]
\[ P(\text{man second} \mid \text{man first}) = \frac{10}{23} \approx 0.435 \]
Different conditions, different answers: the second pick is not independent of the first. That is the whole difference from the binomial. A binomial needs every trial to be a fresh copy of the last one, and here no pick is a fresh copy, because the person picked is gone.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 247 — the softball example
Picture it
Draw both picks as a tree, the tool from Section 3.5. Each branch carries the probability for that pick given everything above it, and each path's probability is the product along it.
Figure (svg): A tree diagram. From the start, a branch of 13/24 leads to W (a woman first) and 11/24 to M. From W, 12/23 leads to W and 11/23 to M; from M, 13/23 leads to W and 10/23 to M. The four ends are labelled with their products over 552: 156, 143, 143 and 110. The two ends where the second pick is a man are blue.
Read the dependence off the picture: the two branches out of W and the two out of M carry different fractions. In a binomial tree every level would repeat the same pair of numbers.
\[ \frac{13}{24} \cdot \frac{11}{23} = \frac{143}{552}, \qquad \frac{11}{24} \cdot \frac{10}{23} = \frac{110}{552} \]
The four end products, 156, 143, 143 and 110 over 552, add to 552 over 552, which is one, as the ends of any complete tree must. The next slide adds up the two blue ends and finds something surprising.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 247 — the dependence of the second pick on the first
Concept
What is the chance that the second person picked is a man, if you do not know who went first? Add the two blue paths of the tree.
\[ P(\text{man second}) = \frac{143}{552} + \frac{110}{552} \]
\[ = \frac{253}{552} = \frac{11}{24} \]
Exactly the chance for the first pick. That is no accident: before anything is revealed, the second person chosen is just as likely to be any of the 24 as the first person is. The same holds for the third pick, the tenth, any pick.
So two statements are true at once, and you need both. Each pick, taken on its own, is a man with probability eleven out of twenty-four. But the picks are dependent: learning who went first changes the chance for the second. The binomial needs independence, so it fails; the mean, which only needs the one-pick chance, will turn out to be easy. Part 4 cashes this in.
This is an addition of our own to the book's section, worked from the book's softball numbers.
Concept
The book lists five characteristics. Read them as a checklist you run on any new problem.
The third characteristic is the cause and the fourth and fifth are its consequences. A Bernoulli trial has the same success probability every time, independently of the others; once items are not put back, neither part holds, as the tree just showed.
hypergeometric random variable — X = the number of items from the group of interest in the sample. Its probabilities form a hypergeometric probability distribution.
The first two characteristics set up the counting: exactly two kinds of item, and one of them singled out by the question. The group of interest is not the larger group or the first one mentioned; it is the one the question asks about.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 247 — the five characteristics
Notation
The book's notation packs the whole setup into three numbers. Read it aloud as: X is a random variable with a hypergeometric distribution, with group sizes r and b and sample size n.
Annotate
On: \( X \sim H(r,\, b,\, n) \)
Order matters. The first parameter is always the group the question counts, so writing the numbers in the wrong order describes a different random variable. Before you write H, write X in words, and the order follows.
The total pool is r plus b. You will need that sum twice later: once in the bottom of the probability formula and once in the mean.
Compare the binomial's notation from Section 4.3, which needed only n and p. The hypergeometric needs three numbers because the chance of success is not fixed: it depends on how many of each group are left, and that depends on both group sizes, not just their ratio.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 249 — notation for the hypergeometric
Worked example
A candy dish contains 100 jelly beans and 80 gumdrops. Fifty candies are picked at random. What is the probability that 35 of the 50 are gumdrops? For now, set the problem up; the number itself comes in Part 3.
Name the two groups
Why: every candy is one or the other.
\[ \text{jelly beans (100)} \quad \text{and} \quad \text{gumdrops (80)} \]
Find the group of interest
Why: the question asks about gumdrops, so gumdrops are the first group.
\[ r = 80 \]
Notice that the group of interest is the smaller group here. Size has nothing to do with it; the wording of the question decides.
The second group and the sample
Why: the other candies, and how many are picked.
\[ b = 100, \qquad n = 50 \]
Define X and write the distribution
Why: X counts gumdrops among the 50.
\[ X \sim H(80,\, 100,\, 50) \]
List the values of X
Why: there are more than 50 of each kind, so every count from 0 to 50 is possible.
\[ x = 0, 1, 2, \ldots, 50 \]
Here the range is the ordinary one, 0 to the sample size, because each kind of candy has more members than the 50 picked: a handful of all jelly beans and a handful of all gumdrops are both possible, however unlikely. Part 2 shows setups where that fails.
Write the probability statement
Why: exactly 35 gumdrops.
\[ P(x = 35) \]
Writing the statement before computing anything is the book's habit throughout this section, and it is a good one. It pins down the event exactly, so that exactly 35, at least 35 and more than 35 cannot be confused later when the arithmetic starts.
Check the setup accounts for every candy
Why: the two groups make up the whole dish, and the sample fits inside it.
\[ r + b = 80 + 100 = 180 \ge n = 50 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.22
Trap
For the candy dish, a first attempt:
\[ r = 100, \quad b = 80 \]
\[ X \sim H(100,\, 80,\, 50) \]
with the reasoning that the first group should be the biggest one, or the one mentioned first.
Wrong. This X counts jelly beans.
The question asks about gumdrops, so gumdrops are the group of interest whatever their size.
\[ r = 80, \quad b = 100 \]
\[ X \sim H(80,\, 100,\, 50) \]
The wrong version is not nonsense: it answers how likely 35 jelly beans are, which is the same event as 15 gumdrops. Always write X in words first, the number of gumdrops in the sample, and read r off that sentence.
The two answers differ enormously: 35 jelly beans is close to typical, while 35 gumdrops is extremely rare. A swapped r does not give a slightly wrong number; it answers a different question.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.22: the group of interest is gumdrops
Worked example
A bag contains letter tiles: 44 vowels and 56 consonants. Seven tiles are picked at random. You want the probability that four of the seven are vowels. Identify the group of interest, its size and the sample size.
Group of interest
Why: the question counts vowels.
\[ \text{vowels}: \quad r = 44 \]
You could equally ask about consonants: four vowels in seven tiles is the same event as three consonants. The book's question is phrased about vowels, so vowels are r. Follow the wording, and your parameters will match anyone else's working and any answer key.
The other group
Why: consonants.
\[ b = 56 \]
The sample
Why: seven tiles, drawn without putting any back.
\[ n = 7 \]
Seven tiles is small next to a bag of 100, and that matters later: when the sample is a small slice of the pool, the dependence between picks is weak. Hold on to that thought for Part 5.
Define X and the question
Why: X counts vowels among the seven tiles.
\[ X \sim H(44,\, 56,\, 7), \qquad P(x = 4) \]
Check the range
Why: with at least seven of each kind, every count from 0 to 7 is possible, and 4 is one of them.
\[ x = 0, 1, \ldots, 7 \quad \text{includes } 4 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Try It 4.22
Worked example
You need a committee of seven students from your organization of 18 women and 15 men, and you care about the number of men. What is the probability that the committee has more than four men? The book asks six setup questions; answer them in order.
(a) With or without replacement?
Why: a person cannot hold two seats on one committee.
\[ \text{without replacement} \]
(b) and (c): the group of interest and its size
Why: the question is about men.
\[ \text{men}: \quad r = 15 \]
The organization is described as 18 women and 15 men, women first and women the larger group. Neither fact matters. You are interested in the number of men, so men are the first group in the book's sense, the one written first inside H.
(d) The other group
Why: the women.
\[ b = 18 \]
(e) Define X and its values
Why: seven seats, and there are at least seven of each group.
\[ X = \text{number of men on the committee}, \quad x = 0, 1, \ldots, 7 \]
Part (f) is where phrasing matters. More than four means five, six or seven; it does not include four. Translate the words into a strict inequality before writing anything else.
(f) The probability statement
Why: more than four, strictly.
\[ P(x > 4) = P(5) + P(6) + P(7) \]
Check the notation against the words
Why: 15 men asked about, 18 women, 7 seats, and the event lists only values in the range.
\[ X \sim H(15,\, 18,\, 7), \quad 5, 6, 7 \le 7 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, pp. 248-249 — Example 4.24
Step zero
A pallet has 200 milk cartons, ten of which have leaked and cannot be sold. A stock clerk randomly chooses 18 for inspection and wants the probability that no more than two are leaking. (The book prints palette; it means a pallet.)
Discussion prompt
Before any calculation: give five reasons this is a hypergeometric problem, one for each characteristic.
The second group is 190, not 200. The parameter b counts only the other group; the total 200 is r plus b. Mixing up the second group with the whole pool is the most common slip in writing the parameters.
Running the five checks takes half a minute and settles the model before any number is written. In the pattern at the end of the deck this is the first box, because everything else depends on it.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 249 — Try It 4.24
Sorting
Section 4.3 gave you the binomial and Section 4.4 the geometric. Sort each experiment into its family. The deciding question is what happens to the pool after each selection.
Sort into buckets
Sort each experiment by the distribution of the count it describes.
The tell for hypergeometric is that the same item cannot be chosen twice: a seat, a card, a laptop. The tell for binomial is a mechanism that resets, a coin or a die, so every trial is a copy of the last.
The laptop item is worth a second look. It sounds like a quality-control binomial, defective or not on each trial, and with a shipment of a million laptops it nearly would be. With a shipment of 100, taking 12 removes an eighth of the pool, and the chances shift noticeably.
The geometric item is different in kind: it has no fixed number of trials at all. When you meet a new problem, ask two questions in order: is the number of trials fixed, and does the pool change after each pick?
Section
Part 2
Concept
A binomial count with n trials can be anything from 0 to n. A hypergeometric count can be cut shorter, because the group of interest might be smaller than the sample. You cannot find more items of a kind than exist.
Two limits apply at once. The count cannot exceed the sample size, since the sample has only n items. And it cannot exceed the size of the group of interest, since only r such items exist.
\[ x \le n \quad \text{and} \quad x \le r \]
\[ \text{largest value} = \min(n,\, r) \]
In the laptop inspection of Example 4.23, the inspector takes 12 laptops but only 10 are defective. Twelve defectives in the sample would need twelve defectives in the shipment. So the largest value is 10, not 12.
\[ \min(12,\, 10) = 10 \]
In the candy dish, by contrast, 80 gumdrops is more than the 50 candies picked, so the sample size is the binding limit and the top value is 50. Always check both numbers.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.23: X may not take on the values 11 or 12
Concept
The bottom of the range can be cut off too, and the book's examples never show it, so it catches people. The sample holds n items; whatever is not from the group of interest must come from the second group, which has only b items.
\[ n - x \le b \quad \Longrightarrow \quad x \ge n - b \]
If n is no bigger than b, that lower limit is zero or negative, and the count can start at 0 as usual. If the sample is bigger than the second group, some seats must be filled from the group of interest.
\[ \text{smallest value} = \max(0,\, n - b) \]
An example of our own: a committee of nine chosen from the six men and five women of Example 4.25. Nine seats, only five women: at least four seats go to men. And only six men exist, so at most six.
Figure (svg): A number line from 0 to 9 counting men on a committee of nine. The values 0 to 3 are crossed out in red, labelled would need 6 or more women; 7 to 9 are crossed out, labelled only 6 men exist. Only 4, 5 and 6 are green and possible.
\[ \max(0,\, 9 - 5) = 4, \qquad \min(9,\, 6) = 6 \]
So X takes only the values 4, 5 and 6. Any committee you imagine with three men would need six women, and there are not six women to choose.
Worked example
A shipment of 100 laptops is known to have ten defective laptops. An inspector randomly chooses 12 for inspection and wants the probability that at most two are defective. Set it up, paying attention to the values of X.
Two groups and the group of interest
Why: the question counts defective laptops.
\[ r = 10 \text{ defective}, \quad b = 90 \text{ non-defective} \]
The sample
Why: twelve laptops, which may be of either kind.
\[ n = 12, \qquad X \sim H(10,\, 90,\, 12) \]
The book is careful to say the twelve may be non-defective or defective. The sample size counts every laptop inspected, not just the ones you are counting; X is the part of the sample that is defective, and n minus X is the part that is fine.
The largest value
Why: only ten defectives exist.
\[ \min(12,\, 10) = 10 \]
The smallest value
Why: 90 good laptops is plenty to fill all twelve places.
\[ \max(0,\, 12 - 90) = 0 \]
Figure (svg): A number line from 0 to 12. Dots mark 0 through 10 as possible values, with 0, 1 and 2 larger and green; 11 and 12 are crossed out in red. A bracket over 0 to 10 reads possible: 0 to 10, only 10 defectives exist; over 11 and 12 it reads impossible. Under 0 to 2: at most two.
So X takes the values 0 through 10, and the values 11 and 12, which a binomial with twelve trials would allow, are impossible. At most two means zero, one or two, the three green values.
Write the probability statement
Why: at most two includes two.
\[ P(x \le 2) = P(0) + P(1) + P(2) \]
Check the values against the book
Why: the book lists 0, 1, 2 and so on up to 10 and rules out 11 and 12, matching both limits.
\[ x \in \{0, 1, \ldots, 10\} \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.23
Worked example
A gross of eggs contains 144 eggs, and a particular gross has 12 cracked ones. An inspector randomly chooses 15 for inspection and wants the probability that at most three are cracked. What is X, and what values does it take?
Define X
Why: the question counts cracked eggs.
\[ X = \text{number of cracked eggs among the 15 inspected} \]
The parameters
Why: 132 uncracked eggs make up the rest of the gross.
\[ r = 12, \quad b = 144 - 12 = 132, \quad n = 15 \]
The problem gives the whole gross, 144, and the cracked eggs inside it, 12. It does not give the second group directly; you get it by subtraction. Writing b as 144 is the slip to avoid: that would count the cracked eggs twice, once in each group.
The same shape as the laptops: the sample, 15, is bigger than the group of interest, 12, so the group size sets the ceiling.
The largest value
Why: only twelve cracked eggs exist.
\[ \min(15,\, 12) = 12 \]
The smallest value
Why: 132 good eggs can fill all fifteen places.
\[ \max(0,\, 15 - 132) = 0 \]
The probability statement
Why: at most three.
\[ X \sim H(12,\, 132,\, 15), \qquad P(x \le 3) \]
Check the parameters add to the gross
Why: the two groups make up the whole box.
\[ r + b = 12 + 132 = 144 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Try It 4.23
Prediction
An intramural basketball team of ten is chosen at random from 15 boys and 12 girls, and you care about the number of boys. The book asks for the group of interest and the sample: boys, 15 of them, and a sample of 10. Now predict the range.
Predict first
What values can X, the number of boys on the team of ten, take?
Correct: 0 through 10
Why: There are 12 girls, enough to fill all ten places, so zero boys is possible. There are 15 boys, more than ten, so the sample size caps the count at 10. Both limits are the ordinary ones here.
\[ \max(0,\, 10 - 12) = 0, \qquad \min(10,\, 15) = 10 \]
The tempting wrong answers come from computing a limit with the wrong group. Three through ten would need only seven girls, and 0 through 15 forgets that a team of ten cannot hold fifteen of anything.
The book stops at naming the group of interest, the boys, and the sample of ten; the probability of eight boys is finished in Part 3. Listing the range first is what tells you eight is a legitimate value to ask about.
Change one number and the answer changes: with only eight girls, ten seats would force at least two boys. The lower limit bites exactly when the other group is smaller than the sample.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 250 — Try It 4.25
Trap
Carrying the habit over from the binomial, for the laptop inspection:
\[ x = 0, 1, 2, \ldots, 12 \]
and then, for P(x more than 8), adding the four terms for 9, 10, 11 and 12 as if every one existed.
Wrong. Two of those values cannot happen.
There are only ten defective laptops, so eleven or twelve in the sample is impossible.
\[ x = 0, 1, \ldots, 10 \]
For a hypergeometric count, compute both ends every time from the group sizes.
\[ \max(0,\, n-b) \le x \]
\[ x \le \min(n,\, r) \]
The formula quietly protects you, since a choose-count of more items than exist is zero, but a table you build by hand will not.
The deeper reason for the trap: in a binomial, each trial can succeed however many successes came before. Here each success uses up one of a limited supply, and once the ten defectives are gone, no further pick can be defective.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.23
Comparison
Fill in the smallest and largest value of X for each setup. Compute both ends from the group sizes: the bottom from the second group, the top from the group of interest and the sample.
Comparison matrix
| setup | r | b | n | smallest x | largest x |
|---|---|---|---|---|---|
| gumdrops (Ex. 4.22) | 80 | 100 | 50 | 0 | 50 |
| defective laptops (Ex. 4.23) | 10 | 90 | 12 | 0 | 10 |
| men, committee of 4 (Ex. 4.25) | 6 | 5 | 4 | 0 | 4 |
| men, committee of 9 | 6 | 5 | 9 | 4 | 6 |
| leaking cartons (Try It 4.24) | 10 | 190 | 18 | 0 | 10 |
| women, softball team of 10 | 13 | 11 | 10 | 0 | 10 |
Only the committee of nine has a lower limit above zero, because it is the only setup where the sample, nine, is bigger than the second group, five. The laptops and the cartons have upper limits below the sample size because the group of interest is smaller than the sample.
A quick way to fill the bottom column: compare the sample size with the second group. If the sample is no bigger than the second group, the bottom is zero. Only when the sample outgrows the second group does the difference become the smallest value.
The softball row shows the ordinary case: thirteen women and eleven men are each enough to fill ten places, so all eleven values from 0 to 10 are possible.
Section
Part 3
Concept
Every hypergeometric probability is a count divided by a count, so first you need a fast way to count choices. How many different pairs can be picked from six men? Order does not matter: choosing Ali then Ben is the same pair as Ben then Ali.
Picking in order gives six choices for the first and five for the second, but each pair has then been counted twice, once in each order. Divide the double counting out.
\[ \binom{6}{2} = \frac{6 \cdot 5}{2 \cdot 1} = 15 \]
Figure (svg): Left: six blue points M1 to M6 at the corners of a hexagon with all 15 connecting lines drawn. Middle: five orange points W1 to W5 at the corners of a pentagon with all 10 connecting lines. Right: the text choose 2 of 6 men, 15 ways; choose 2 of 5 women, 10 ways; one line from each side gives 15 times 10 = 150 committees with exactly 2 men.
The picture checks the count: fifteen lines join six points. The same move for the five women gives ten pairs, and the pentagon shows ten lines.
\[ \binom{5}{2} = \frac{5 \cdot 4}{2 \cdot 1} = 10 \]
In general, the number of ways to choose k items from n, ignoring order, is n choose k. On a TI calculator it is the nCr key; in a spreadsheet it is COMBIN.
\[ \binom{n}{k} = \frac{n!}{k!\,(n-k)!} \]
The combination count is not new in the book; the binomial formula of Section 4.3 uses it too. This section's book text leaves the counting to a computer, so this slide supplies it.
Concept
Choosing a committee at random means every possible committee has the same chance. When outcomes are equally likely, a probability is the number of favourable outcomes over the number of possible outcomes, as in Chapter 3.
Count the possible committees of four from the eleven people of Example 4.25.
\[ \binom{11}{4} = \frac{11 \cdot 10 \cdot 9 \cdot 8}{4 \cdot 3 \cdot 2 \cdot 1} = \frac{7920}{24} = 330 \]
Now count the committees with exactly two men. Such a committee is two men and two women, and you can pick the men and the women separately: any pair of men goes with any pair of women.
\[ \binom{6}{2}\binom{5}{2} = 15 \cdot 10 = 150 \]
So the probability is 150 favourable committees out of 330. Replace the numbers by the parameters and you have the formula for the whole section: choose x from the group of interest, choose the rest of the sample from the other group, and divide by all the ways to choose the sample.
\[ P(x) = \frac{\dbinom{r}{x}\dbinom{b}{n-x}}{\dbinom{r+b}{n}} \]
The book computes this with a calculator or computer and prints only the answer. The formula is what that software is doing.
Notation
Three combination counts, each answering one plain question. Read the formula as: ways to get x from the group of interest, times ways to fill the rest from the other group, over all possible samples.
Annotate
On: \( P(x) = \frac{\dbinom{r}{x}\dbinom{b}{n-x}}{\dbinom{r+b}{n}} \)
Two bookkeeping checks are built into the formula, and they catch most slips. The top numbers of the two upper counts add to the top number below: r plus b. The bottom numbers of the two upper counts add to the bottom number below: x plus n minus x is n.
If either sum fails, a group has been miscounted or the sample has not been fully accounted for. Do this check before any arithmetic.
Worked example
A school site committee of four is chosen at random from six men and five women. What is the probability that two of them are men? The book gives only the answer, 0.4545, from a computer; here is the count behind it.
Set up the distribution
Why: X counts men, the group of interest.
\[ X \sim H(6,\, 5,\, 4), \qquad x = 0, 1, 2, 3, 4 \]
Run the bookkeeping check
Why: tops 6 and 5 make 11; bottoms 2 and 2 make 4.
\[ P(x = 2) = \frac{\dbinom{6}{2}\dbinom{5}{2}}{\dbinom{11}{4}} \]
Count the pairs of men
Why: six choose two.
\[ \binom{6}{2} = 15 \]
Count the pairs of women
Why: the other two seats.
\[ \binom{5}{2} = 10 \]
Count all committees
Why: eleven choose four.
\[ \binom{11}{4} = 330 \]
Keep the three counts apart until the end. The top two describe one kind of committee, two men and two women; the bottom one describes every committee there could be. Mixing them up, for instance dividing by 11 choose 2, is how a probability above one appears.
Divide
Why: favourable committees over all committees.
\[ P(x = 2) = \frac{15 \cdot 10}{330} = \frac{150}{330} \approx 0.4545 \]
In context: the probability that exactly two of the four committee members are men is about 0.45, a little under one half. That matches the book's 0.4545. For a spreadsheet, HYPGEOM.DIST(2, 4, 6, 11, FALSE) returns the same value.
Check by picking one seat at a time
Why: one order, man, man, woman, woman, has the product below; each of the six orders of two men and two women gives the same product.
\[ 6 \cdot \frac{6 \cdot 5 \cdot 5 \cdot 4}{11 \cdot 10 \cdot 9 \cdot 8} = 6 \cdot \frac{600}{7920} \approx 0.4545 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 249 — Example 4.25
Concept
The check on the last slide is worth understanding, because it shows the formula and the multiplication rule are the same idea. Take one order of picks: man, man, woman, woman.
\[ \frac{6}{11} \cdot \frac{5}{10} \cdot \frac{5}{9} \cdot \frac{4}{8} = \frac{600}{7920} \]
Each fraction shrinks the pool by one, because nothing goes back. Now take a different order, woman, man, woman, man. The fractions come in a different order but they are the same fractions, so the product is the same.
\[ \frac{5}{11} \cdot \frac{6}{10} \cdot \frac{4}{9} \cdot \frac{5}{8} = \frac{600}{7920} \]
How many orders are there? Choose which two of the four seats hold the men: four choose two, which is six.
\[ \binom{4}{2} \cdot \frac{600}{7920} = \frac{3600}{7920} = \frac{150}{330} \approx 0.4545 \]
So the combination formula is the multiplication rule done in bulk. Use whichever route you find clearer; for anything bigger than a small committee, the formula saves writing out long chains of fractions.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 247 — the dependent picks of characteristic four
Picture it
Now count every committee, not just those with two men. Each committee has between zero and four men, so the 330 committees fall into five piles. Lay them end to end.
Figure (svg): One long bar representing all 330 committees, cut into five pieces in proportion: 5 committees with no men, 60 with one man, 150 with two, 100 with three and 15 with four. Under each piece is its share of 330.
Each pile is counted exactly as before: choose the men, choose the women, multiply. The piles must add back to all 330 committees, because every committee lands in exactly one pile.
\[ 5 + 60 + 150 + 100 + 15 = 330 \]
Because every committee is equally likely, the probability of each value of X is simply its share of the bar. That is the whole distribution, drawn: two men is the most common by a wide margin, and an all-women committee happens only five times in 330.
Notice the lopsidedness. There is one more man than woman in the pool, and it shows: three men, 100 committees, is noticeably more common than one man, 60.
Worked example
Complete the table for every value of X, then draw it. The value x = 2 is done; the others follow the same three counts.
x = 0: no men, four women
Why: one way to choose no men.
\[ \binom{6}{0}\binom{5}{4} = 1 \cdot 5 = 5 \]
Choosing no men can be done in exactly one way, by leaving them all out, so six choose zero is 1. An all-women committee is then just a choice of which four of the five women serve, which is the same as choosing the one woman left out: five ways.
x = 1: one man, three women
Why: six men to choose from, ten triples of women.
\[ \binom{6}{1}\binom{5}{3} = 6 \cdot 10 = 60 \]
x = 3: three men, one woman
Why: twenty triples of men, five single women.
\[ \binom{6}{3}\binom{5}{1} = 20 \cdot 5 = 100 \]
x = 4: four men, no women
Why: fifteen sets of four men.
\[ \binom{6}{4}\binom{5}{0} = 15 \cdot 1 = 15 \]
Notice the symmetry in the counts: six choose four equals six choose two, since picking four men to serve is the same as picking two to leave out. Symmetries like that are a quick way to catch an arithmetic slip in a row.
| x | 0 | 1 | 2 | 3 | 4 |
|---|---|---|---|---|---|
| committees | 5 | 60 | 150 | 100 | 15 |
| P(x) | 0.0152 | 0.1818 | 0.4545 | 0.3030 | 0.0455 |
Figure (svg): A probability histogram of the number of men on a committee of four: bars at 0 to 4 with heights 5/330, 60/330, 150/330, 100/330 and 15/330. A yellow triangle under the baseline marks the mean at 24/11, about 2.18.
This is the book's Figure 4.5 redrawn from the counts. The histogram peaks at 2 men and leans slightly right, because men are a little over half the pool.
Check the distribution is complete
Why: the counts must add to every committee, so the probabilities add to one.
\[ \frac{5 + 60 + 150 + 100 + 15}{330} = \frac{330}{330} = 1 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 250 — Figure 4.5
Error analysis
Four students computed the probability of exactly two men on the committee of four. Only one is right. Find what each of the others did.
\[ \text{(i) } \frac{15 \cdot 10}{330} = 0.4545 \qquad \text{(ii) } \frac{15}{330} = 0.0455 \]
\[ \text{(iii) } \frac{15 \cdot 5}{330} = 0.2273 \qquad \text{(iv) } 6\left(\tfrac{6}{11}\right)^{2}\left(\tfrac{5}{11}\right)^{2} = 0.3688 \]
Annotate
On: \( 0.4545 \quad | \quad 0.0455 \quad | \quad 0.2273 \quad | \quad 0.3688 \)
Two of the errors are caught by the bookkeeping check from the notation slide: in the second the bottoms add to two, and in the third they add to six, never the four seats of the committee.
A second quick test: compare with the picture of all 330 committees. Two men owned almost half the bar, so any answer far from one half should make you suspicious before you look for the specific mistake.
The binomial attempt is the most tempting, because it is last section's formula and gives a sensible-looking number. Part 5 measures exactly how far off it is here, and when it would be close enough.
Fill the middle
The basketball team of ten from 15 boys and 12 girls: find the probability that exactly eight players are boys. The formula is set up; fill in the missing pieces.
\[ P(x = 8) = \frac{\dbinom{15}{8}\dbinom{12}{2}}{\dbinom{27}{10}} = \frac{6435 \cdot 66}{8{,}436{,}285} \]
Fill in the blanks
n - x = 10 - 8 = 2, \quad \binom660.0503 = ___, \quad P(x = 8) \approx ___
Why: Eight boys leave two seats for girls, and there are 12 times 11 over 2, which is 66, ways to choose two girls. The numerator is 6435 times 66, which is 424,710, and dividing by 8,436,285 gives about 0.0503.
Eight boys out of ten is unlikely, about one team in twenty, because boys are only 15 of the 27 players. The expected number of boys is ten times fifteen twenty-sevenths, about 5.6, and eight is well above that.
Run the bookkeeping check on the setup line: the tops 15 and 12 make 27, and the bottoms 8 and 2 make 10. If you wrote 12 choose 8 for the girls, the bottoms would add to 16.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 250 — Try It 4.25
Worked example
Back to the letter tiles: 44 vowels, 56 consonants, seven tiles drawn. You set it up in Part 1 as X following H(44, 56, 7). Now compute the probability that exactly four are vowels.
Write the formula and check it
Why: tops 44 and 56 make 100, bottoms 4 and 3 make 7.
\[ P(x = 4) = \frac{\dbinom{44}{4}\dbinom{56}{3}}{\dbinom{100}{7}} \]
Choose four vowels
Why: 44 times 43 times 42 times 41, over 4 factorial.
\[ \binom{44}{4} = 135{,}751 \]
Choose three consonants
Why: 56 times 55 times 54, over 3 factorial.
\[ \binom{56}{3} = 27{,}720 \]
Each count ignores the order within its own group, because a hand of tiles is a set: drawing A then E is the same four vowels as E then A. That is why each product of falling numbers is divided by a factorial, which removes the orderings.
Multiply
Why: favourable samples.
\[ 135{,}751 \cdot 27{,}720 = 3{,}763{,}017{,}720 \]
Count all samples of seven
Why: a large number, which is why software is used.
\[ \binom{100}{7} = 16{,}007{,}560{,}800 \]
Divide
Why: favourable over total.
\[ P(x = 4) = \frac{3{,}763{,}017{,}720}{16{,}007{,}560{,}800} \approx 0.2351 \]
Here the sample is only 7 of 100 tiles, so the dependence between draws is mild; a binomial with p equal to 0.44 would give about 0.2304, close to the exact answer. Part 5 explains when that closeness can be trusted.
In context: about 24 percent of the time, a draw of seven tiles holds exactly four vowels. With vowels making up 44 percent of the bag, the expected count is about 3.1, so four is a common outcome but not the most common one.
Check by the one-order route
Why: one order of four vowels then three consonants, times the 35 ways to place the vowels among seven draws.
\[ 35 \cdot \frac{44 \cdot 43 \cdot 42 \cdot 41 \cdot 56 \cdot 55 \cdot 54}{100 \cdot 99 \cdots 94} \approx 35(0.0067165) \approx 0.2351 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Try It 4.22
Worked example
The committee of seven from 15 men and 18 women. In Part 1 you reduced the question to three terms. Now count them.
All committees of seven
Why: thirty-three choose seven.
\[ \binom{33}{7} = 4{,}272{,}048 \]
Over four million possible committees, far too many to list. That is the point of counting: you never write the committees down, you only count how many of each kind there are. Each of the next three lines counts one kind.
Five men and two women
Why: check: 5 plus 2 is 7.
\[ \binom{15}{5}\binom{18}{2} = 3003 \cdot 153 = 459{,}459 \]
Six men and one woman
Why: check: 6 plus 1 is 7.
\[ \binom{15}{6}\binom{18}{1} = 5005 \cdot 18 = 90{,}090 \]
Seven men and no women
Why: every seat to a man.
\[ \binom{15}{7}\binom{18}{0} = 6435 \cdot 1 = 6435 \]
You can add the three counts before dividing because the three kinds of committee cannot overlap: no committee has both five and six men. Adding counts of separate kinds is the addition rule for mutually exclusive events from Chapter 3, done with whole numbers.
Add the favourable counts and divide
Why: one fraction for the whole event.
\[ P(x > 4) = \frac{459{,}459 + 90{,}090 + 6435}{4{,}272{,}048} = \frac{555{,}984}{4{,}272{,}048} \approx 0.1301 \]
Figure (svg): A probability histogram of the number of men on a committee of seven from 15 men and 18 women, x = 0 to 7. Bars 0 to 4 are blue: 0.007, 0.065, 0.211, 0.326, 0.261. Bars 5 to 7 are yellow: 0.108, 0.021, 0.002. The mean, about 3.18, is marked.
In context: there is about a 13 percent chance that a randomly chosen committee of seven has more than four men. The three yellow bars hold that probability; most of it sits on exactly five.
Check with the other five piles
Why: the committees with 0 to 4 men number 3,716,064, and both parts together must be every committee.
\[ 3{,}716{,}064 + 555{,}984 = 4{,}272{,}048 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, pp. 248-249 — Example 4.24
Worked example
The inspector takes 12 laptops from 100, of which 10 are defective, and wants the probability of at most two defective. The counts are enormous, but the steps are the same three.
All samples of twelve
Why: one hundred choose twelve.
\[ \binom{100}{12} = 1{,}050{,}421{,}051{,}106{,}700 \]
No defectives
Why: all twelve from the 90 good laptops.
\[ \binom{10}{0}\binom{90}{12} = 273{,}897{,}571{,}557{,}780 \]
Numbers this size are why the book hands the calculation to a computer. You do not need to carry fifteen digits by hand; what you need is to know which three counts the software is dividing, so that you can set it up and judge whether its answer is sensible.
One defective
Why: one of ten, eleven of ninety.
\[ \binom{10}{1}\binom{90}{11} = 416{,}046{,}944{,}138{,}400 \]
Two defectives
Why: two of ten, ten of ninety.
\[ \binom{10}{2}\binom{90}{10} = 257{,}429{,}046{,}685{,}635 \]
Divide each count by the total
Why: the three probabilities.
\[ P(0) \approx 0.2608, \quad P(1) \approx 0.3961, \quad P(2) \approx 0.2451 \]
Add
Why: at most two is the three values together.
\[ P(x \le 2) \approx 0.2608 + 0.3961 + 0.2451 = 0.9019 \]
Figure (svg): A probability histogram of the number of defective laptops among 12 inspected, x = 0 to 6. The bars at 0, 1 and 2 are green, with heights 0.2608, 0.3961 and 0.2451; the bars at 3 to 6 are grey and small: 0.0807, 0.0155, 0.0018 and 0.0001. The mean, 1.2, is marked with a triangle.
In context: if the inspector samples twelve laptops from this shipment, there is about a 90 percent chance of finding two or fewer defective ones. An inspection rule that rejects the shipment only on three or more defectives would pass this bad shipment nine times in ten.
Check the first term one pick at a time
Why: no defectives means twelve good laptops in a row, each fraction one smaller.
\[ \frac{90}{100} \cdot \frac{89}{99} \cdots \frac{79}{89} \approx 0.2608 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.23
Worked example
The gross of 144 eggs with 12 cracked, 15 inspected: X follows H(12, 132, 15), and you want the probability of at most three cracked. Here the counts run to twenty digits, so the probabilities are shown rounded.
The shape of every term
Why: cracked from 12, uncracked from 132, fifteen in all.
\[ P(x) = \frac{\dbinom{12}{x}\dbinom{132}{15-x}}{\dbinom{144}{15}} \]
Compare this with the laptops. There the inspector took 12 from 100 with 10 bad; here 15 from 144 with 12 bad. The share of bad items is similar, about one in ten or twelve, so you should expect a similar picture: most of the probability on zero, one and two.
The four terms
Why: x = 0, 1, 2, 3, each computed exactly and then rounded.
\[ 0.2525, \quad 0.3851, \quad 0.2492, \quad 0.0900 \]
Add them
Why: the unrounded terms add to 0.97673; the rounded ones add to 0.9768 because of rounding.
\[ P(x \le 3) \approx 0.9767 \]
Figure (svg): A probability histogram of cracked eggs among 15 inspected from a gross with 12 cracked, x = 0 to 6. Bars 0 to 3 are blue with heights 0.2525, 0.3851, 0.2492 and 0.0900; bars 4 to 6 are red and small: 0.0201, 0.0029 and 0.0003. The mean, 1.25, is marked.
In context: an inspector checking fifteen eggs from this gross finds three or fewer cracked about 98 percent of the time. The expected number found is fifteen times twelve over 144, which is 1.25 eggs.
Check with the complement
Why: the red bars, x = 4 to 12, must hold the rest.
\[ P(4) + P(5) + \cdots + P(12) \approx 0.0233 = 1 - 0.9767 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Try It 4.23
Estimation
The candy dish: 80 gumdrops and 100 jelly beans, 50 picked. The book stops at writing P(x = 35). Estimate first: gumdrops are 80 of 180 candies, about 44 percent, so a typical handful of 50 holds about 22 of them.
Predict first
How likely is it that 35 of the 50 candies are gumdrops?
Correct: far less than 1 in 1,000
Why: Thirty-five is thirteen above the typical 22, and in a sample of 50 that is a very long way. The exact value is about 0.0000144, roughly 1 in 70,000.
\[ P(x = 35) = \frac{\dbinom{80}{35}\dbinom{100}{15}}{\dbinom{180}{50}} \approx 0.0000144 \]
Figure (svg): A probability histogram of the number of gumdrops among 50 candies picked from 80 gumdrops and 100 jelly beans, for x from 8 to 38. The bars form a hump peaking at 22 with height about 0.13; a yellow triangle marks the mean at about 22.2. An arrow points to x = 35, where the bar is too small to see: P is about 0.0000144.
The picture says it at once: the probability lives in a hump around 22, and 35 is out in the tail where the bars vanish. Checking where a value sits against the mean is a quick sense test for any answer you compute.
The last option is a common confusion: 35 out of 50 is a sample proportion, not a probability. The probability is about how often a handful looks like that, and the answer is almost never.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.22
Matching
Match each piece of the formula, for the committee of four from six men and five women, to what it counts.
Match the pairs
Why: Six choose two picks the men, five choose two fills the remaining seats with women, and eleven choose four counts every possible committee. Favourable committees, 15 times 10, over all committees gives the probability.
Say each piece in words whenever you set up a problem: ways to choose the ones I am counting, ways to fill the rest, all the ways. A formula you can say is a formula you will not misassemble.
Notice that the fourth item is the only one that is a probability. The first three are counts of committees, whole numbers, and only the ratio of favourable to total becomes a number between zero and one.
The second piece is the one most often dropped. Choosing the men alone does not make a committee; each choice of men has to be completed by some choice of women, and there are ten ways to do it.
Section
Part 4
Concept
Part 1 found that each softball pick, taken on its own, has the same chance of being a man as the first pick. The same holds for any hypergeometric experiment: each of the n picks is from the group of interest with probability r over r plus b.
Figure (svg): Four boxes labelled seat 1 to seat 4, each saying man: 6/11, joined by plus signs. Below: expected men = 6/11 + 6/11 + 6/11 + 6/11 = 24/11, about 2.18.
An expected count adds up the chances seat by seat, and this works even though the seats are dependent. Expected values always add, whether or not the pieces are independent.
\[ \mu = n \cdot \frac{r}{r+b} = \frac{nr}{r+b} \]
Read it as the sample size times the proportion of the pool that belongs to the group of interest. Four seats and six elevenths of the pool are men, so expect four times six elevenths.
\[ \mu = \frac{(4)(6)}{6+5} = \frac{24}{11} \approx 2.18 \]
This is the book's formula, and the same mean a binomial with p equal to r over r plus b would give. The book writes the answer as m = 2.18; the m is a lost mu.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 250 — the formula for the mean
Worked example
The second question of Example 4.25: how many men do you expect on the committee of four? Compute it with the formula, then confirm it from the distribution table.
Identify n and the proportion of men
Why: four seats, six men in a pool of eleven.
\[ n = 4, \qquad \frac{r}{r+b} = \frac{6}{11} \]
Apply the formula
Why: sample size times proportion.
\[ \mu = 4 \cdot \frac{6}{11} = \frac{24}{11} \]
Convert to a decimal
Why: as the book does.
\[ \mu \approx 2.18 \]
In context: you would expect about 2.18 men, that is about two, on a randomly chosen committee of four. No committee has 2.18 men; it is the long-run average over many committees chosen this way.
A formula you have only been told deserves a check. Section 4.2 gave you a method that works for any distribution, value times probability summed, and you already have the whole table from Part 3. If the shortcut is right, the two must agree exactly.
Build the expected value from the table
Why: Section 4.2's method: value times probability, using the counts over 330.
\[ \mu = \frac{0(5) + 1(60) + 2(150) + 3(100) + 4(15)}{330} \]
Add the numerators
Why: 0, 60, 300, 300 and 60.
\[ \mu = \frac{720}{330} \]
Check the two routes agree
Why: 720 over 330 reduces to 24 over 11.
\[ \frac{720}{330} = \frac{24}{11} \approx 2.18 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, pp. 249-250 — Example 4.25
Comparison
Apply the mean formula to each setup: the sample size times the share of the pool in the group of interest. Fill in the blanks, to two decimal places where needed.
Comparison matrix
| setup | n | r/(r + b) | μ = nr/(r + b) |
|---|---|---|---|
| gumdrops (Ex. 4.22) | 50 | 80/180 | 22.22 |
| defective laptops (Ex. 4.23) | 12 | 10/100 | 1.2 |
| men, committee of 7 (Ex. 4.24) | 7 | 15/33 | 3.18 |
| leaking cartons (Try It 4.24) | 18 | 10/200 | 0.9 |
| vowels (Try It 4.22) | 7 | 44/100 | 3.08 |
Each answer is where the distribution balances, and you can use it to sense-check a probability. The gumdrop mean of about 22 is why 35 gumdrops was so unlikely; the laptop mean of 1.2 is why at most two defectives was so likely.
Every mean here is a sample size times a share, never a sample size times a count. Say the share aloud as a fraction of the whole pool, ten out of a hundred laptops, fifteen out of thirty-three club members, and the denominator cannot go wrong.
For the cartons, the share of leaking cartons is 10 out of 200, one in twenty, so eighteen cartons hold 0.9 leaking cartons on average. The clerk will often find none at all.
Trap
For the committee of four from six men and five women:
\[ \mu = \frac{nr}{b} = \frac{4 \cdot 6}{5} \]
\[ = 4.8 \text{ men} \]
Wrong, and visibly so: a committee of four cannot average 4.8 men.
The fraction must be the share of the whole pool in the group of interest, so its bottom is the whole pool.
\[ \mu = \frac{nr}{r+b} = \frac{24}{11} \]
\[ \approx 2.18 \]
A quick range test catches the slip: a mean must lie between the smallest and largest value of X, here 0 and 4. Anything outside that range is an arithmetic error, not an answer.
The slip is subtle when the numbers are large. For the candy dish, fifty times eighty over one hundred gives 40, which is inside the range and looks fine; only the share check, 80 out of 180, gives the right 22.2.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 250 — the formula for the mean
Prediction
The hypergeometric committee and a binomial with p equal to six elevenths have the same mean, 2.18 men. Think about the extremes: an all-male committee needs four men in a row from a pool that loses a man with every pick.
Predict first
Compared with the binomial with the same mean, the hypergeometric distribution of men on the committee is
Correct: less spread out
Why: Without replacement, each man chosen makes the next man less likely and each woman chosen makes the next man more likely. The pool pushes back against runs, so extreme committees are rarer and the count stays closer to its mean.
\[ \text{hypergeometric: } P(0) + P(4) \approx 0.0152 + 0.0455 = 0.0606 \]
\[ \text{binomial: } P(0) + P(4) \approx 0.0427 + 0.0885 = 0.1312 \]
The two extreme values together are about twice as likely under the binomial as under the hypergeometric. Sampling without replacement is self-correcting: an unusual run uses up that group and makes the next pick lean the other way.
You can feel this with a tiny pool. Draw two people from one man and one woman: without replacement you always get exactly one man, with no spread at all, while with replacement you could get zero, one or two. Removing items squeezes the count toward its mean. The next slide puts a number on it.
Concept
This slide goes beyond the book's section, which gives only the mean. Write p for the proportion in the group of interest and N for the pool size r plus b. The hypergeometric variance is the binomial variance times a correction factor.
\[ \sigma^2 = n p (1-p) \cdot \frac{N-n}{N-1} \]
The factor is less than one whenever you sample more than one item, and it measures how much the shrinking pool pulls the count toward its mean. For the committee, p is six elevenths, N is 11 and n is 4.
\[ \sigma^2 = 4 \cdot \frac{6}{11} \cdot \frac{5}{11} \cdot \frac{7}{10} = \frac{84}{121} \approx 0.6942 \]
\[ \sigma \approx 0.8332 \qquad \text{binomial: } \sqrt{4 \cdot \tfrac{6}{11} \cdot \tfrac{5}{11}} \approx 0.9959 \]
Verify it with Section 4.2's shortcut on the distribution table: the average of x squared is 1800 over 330, and subtracting the mean squared gives the same variance.
\[ \frac{1800}{330} - \left(\frac{24}{11}\right)^2 \approx 5.4545 - 4.7603 = 0.6942 \]
When the pool is huge and the sample small, N minus n over N minus 1 is almost exactly one, and the two spreads agree. That is the next part in one line.
Section
Part 5
Concept
Sampling without replacement always makes the picks dependent. But how dependent depends on the size of the pool. Suppose a city of 50 million people, 30 million of whom support a policy, and a pollster samples 1,000 of them.
\[ P(\text{first supports}) = \frac{30{,}000{,}000}{50{,}000{,}000} = 0.6 \]
\[ P(\text{second supports} \mid \text{first supports}) = \frac{29{,}999{,}999}{49{,}999{,}999} \approx 0.599999992 \]
The second chance differs from the first in the ninth decimal place. Even after a thousand picks the pool has lost only two thousandths of one percent of its people, so every pick has a chance within a whisker of 0.6.
\[ \frac{n}{N} = \frac{1000}{50{,}000{,}000} = 0.00002 \]
The picks are still technically dependent, but a binomial with p equal to 0.6 describes them almost perfectly. This is why polls are analysed with binomial methods even though nobody is asked twice.
With the committee of four from eleven, the sampling fraction is four elevenths, over a third of the pool, and the dependence is strong. The size of that fraction is the whole story.
Worked example
How wrong is the binomial for the committee of four from six men and five women? Compute the binomial answer for exactly two men and set it against the correct 0.4545.
The binomial parameters
Why: four trials, each a man with the pool's proportion.
\[ n = 4, \qquad p = \frac{6}{11}, \qquad q = \frac{5}{11} \]
This is the model the error-analysis slide warned about. It is not a silly model: p is the right proportion of men and the mean it gives is correct. What it gets wrong is the dependence, and you are about to see how much that costs.
The binomial formula
Why: Section 4.3.
\[ P(x = 2) = \binom{4}{2}\left(\frac{6}{11}\right)^2\left(\frac{5}{11}\right)^2 \]
Multiply out
Why: six orders, each with probability 36 times 25 over 14,641.
\[ = \frac{6 \cdot 36 \cdot 25}{14{,}641} = \frac{5400}{14{,}641} \]
Divide
Why: a decimal to compare.
\[ \approx 0.3688 \]
The binomial answer, 0.3688, is about a fifth too small. It pretends each chosen person goes back into the pool, which lets the same man be picked twice and makes lopsided committees more likely than they really are.
Check the size of the error
Why: the difference against the correct value, relative to it.
\[ \frac{0.4545 - 0.3688}{0.4545} \approx 0.19 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 249 — Example 4.25, for the correct value
Picture it
Keep the committee of four and the proportion of men at six elevenths, but grow the pool from 11 people to 1,100: 600 men and 500 women. Draw the hypergeometric and the binomial side by side for both pools.
Figure (svg): Two panels of paired bars for x = 0 to 4 men on a committee of four; blue bars are hypergeometric and orange bars binomial with p = 6/11. Left, with 11 people, the blue bar at 2 is much taller (0.4545 against 0.3688) and the blue bars at 0 and 4 are much shorter. Right, with 1,100 people in the same proportion, each blue bar is almost exactly as tall as its orange partner.
On the left, with eleven people, the bars disagree at every value: the hypergeometric is taller in the middle and shorter at both ends, the smaller spread you predicted in Part 4.
On the right, with 1,100 people, each blue bar almost matches its orange partner. The sample of four is now well under one percent of the pool, and removing four people from 1,100 hardly changes anyone's chances.
\[ P_{H}(2): \; 0.4545 \to 0.3695, \qquad P_{B}(2) = 0.3688 \]
Look at the ends especially. With eleven people, an all-male committee needs four of the six men in a row, and each man chosen leaves fewer; the hypergeometric bar at four is half the binomial's. With 1,100 people, four men in a row barely dents the 600, and the bars match.
The binomial did not change between the panels; it only depends on n and p. What changed is the hypergeometric, which approaches the binomial as the pool grows around a fixed sample.
Concept
How big must the pool be? A widely used rule of thumb, not stated in this section of the book, compares the sample size with the population size.
\[ \frac{n}{N} < 0.05 \quad \Longrightarrow \quad \text{a binomial with } p = \frac{r}{r+b} \text{ is acceptable} \]
Why 5 percent? At that sampling fraction the correction factor on the spread is still about 0.975, so the binomial overstates the standard deviation by only about two and a half percent, and the probabilities agree to about two decimal places.
\[ n = 10, \; N = 200: \quad \sqrt{\frac{190}{199}} \approx 0.977 \]
Run the section's examples through it. The committee of four from eleven samples 36 percent of the pool, far over. The laptops sample 12 percent, over. The eggs sample about 10 percent, over. The tiles sample 7 percent, just over. A poll of 1,000 from 50 million samples 0.002 percent, far under.
The rule is a convenience, not a law. The hypergeometric is always the exact model when nothing is put back; the binomial is a shortcut you are allowed when it cannot mislead you.
Tweak it
The graph shows the factor that multiplies the binomial standard deviation to give the hypergeometric one, the square root of N minus n over N minus 1, as the population N grows. The slider sets the sample size.
Parameter explorer
Set the sample size to 10. Read the factor at a population of 200, where the sample is 5 percent of it. Then set the sample to 20 and find roughly where the factor reaches the same value.
\[ \text{factor} = \sqrt{\frac{N - {n}}{N - 1}} \]
\[ n = 10, \; N = 200: \; 0.977 \qquad n = 20, \; N = 400: \; \sqrt{\tfrac{380}{399}} \approx 0.976 \]
Doubling the sample doubles the population needed for the same accuracy: the factor depends almost entirely on the fraction n over N. That is why the rule of thumb is written as a fraction and not as a fixed population size.
Slide n down to 1. The factor is exactly one for every population: a single pick has nothing to depend on, so one draw without replacement is the same as one draw with it.
Worked example
The laptop inspection samples 12 of 100, a sampling fraction of 12 percent, over the five-percent line. See how much the binomial misses the correct 0.9019.
The binomial parameters
Why: twelve laptops, each defective with the shipment's proportion.
\[ n = 12, \qquad p = \frac{10}{100} = 0.1 \]
No defectives
Why: every laptop good.
\[ P(0) = 0.9^{12} \approx 0.2824 \]
One defective
Why: twelve positions for it.
\[ P(1) = 12(0.1)(0.9)^{11} \approx 0.3766 \]
Two defectives
Why: sixty-six pairs of positions.
\[ P(2) = 66(0.1)^2(0.9)^{10} \approx 0.2301 \]
Set these against the exact terms one at a time: 0.2824 against 0.2608 for none, 0.3766 against 0.3961 for one, 0.2301 against 0.2451 for two. The binomial puts too much on zero and too little on one and two, because it lets the pool of good laptops refill.
Add
Why: the binomial at most two.
\[ P_{B}(x \le 2) \approx 0.8891 \qquad \text{against} \qquad P_{H}(x \le 2) \approx 0.9019 \]
The binomial is off by about 0.013, small but real, and in the direction the spread argument predicts: the binomial has fatter tails, so it puts too much probability on three or more defectives. An inspector who used it would overstate the chance of catching this shipment.
Check against the rule of thumb
Why: 12 percent is over 5 percent, so a visible gap is what the rule leads you to expect.
\[ \frac{12}{100} = 0.12 > 0.05 \]
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, p. 248 — Example 4.23, for the setup
Discrimination
Each situation samples without replacement. Compute the sampling fraction in your head and sort by whether the five-percent rule allows a binomial.
Sort into buckets
Sort each situation by whether a binomial approximation is acceptable under the five-percent rule.
Card hands are the classic surprise. Five cards feel like a small sample, but the deck has only 52, so each card dealt shifts the chances for the next by a visible amount. A hand of five hearts is hypergeometric, and the binomial answer would be noticeably off.
The candy dish is the other surprise: 180 candies feels like a lot, but taking 50 of them removes more than a quarter, and after thirty picks the share of gumdrops left can be quite different from where it started.
Notice that the kind of object never decides it, only the fraction. Thirty bolts from a lot of ten thousand is fine; four people from a club of eleven is not.
Counterexample
A classmate states a rule: problems about cards are always hypergeometric, and problems about coins and dice are always binomial.
Discussion prompt
Find a card problem that is binomial, and explain what makes the difference.
That rule works for the usual wording because cards are usually dealt and coins are always re-tossed. But the rule is about habit, not mechanism, and exam writers like to break it by adding the words with replacement.
A second counterexample runs the other way in spirit: drawing five cards from a shoe of eight shuffled decks is still without replacement, so strictly hypergeometric, but five cards from 416 is a sampling fraction near one percent, and a binomial is fine. Cards, again, do not decide it.
The reverse is impossible: a coin toss cannot be hypergeometric, because there is no finite pool of heads for the tosses to use up. The five characteristics need two groups of items that can run out.
Real world
Many national lotteries draw six balls from 49 without replacement, and you win prizes according to how many of your six numbers come up. That is a hypergeometric experiment: your six numbers are the group of interest, the other 43 are the second group, and six balls are drawn.
Discussion prompt
Write the distribution of X, the number of your numbers drawn. Find the probability of matching all six, the probability of matching exactly three, and the expected number of matches.
The sampling fraction is six out of 49, about 12 percent, so the binomial would not do here. Lottery operators use exactly this formula to set the prize for each number of matches.
Check the bookkeeping on the three-match line: the tops 6 and 43 make 49 balls, and the bottoms 3 and 3 make the 6 balls drawn. Your numbers are the group of interest because the question counts matches, not because they are special balls.
Combine it with Section 4.2: the expected value of a ticket is the sum of each prize times its hypergeometric probability, minus the price, and it is always negative.
Comparison
Sections 4.3 to 4.5 each built a distribution from a different sampling story. Fill in the blanks to line them up.
Comparison matrix
| binomial (4.3) | geometric (4.4) | hypergeometric (4.5) | |
|---|---|---|---|
| X counts | successes in n trials | trials until the first success | items from the group of interest |
| trials independent? | yes | yes | no |
| number of trials | fixed, n | not fixed | fixed, n |
| parameters | n, p | p | r, b, n |
| mean | np | 1/p | nr/(r + b) |
The binomial and hypergeometric rows match everywhere except independence, and their means are the same formula once p is written as r over r plus b. That is why one approximates the other when the pool is large.
The geometric is the odd one out: it counts trials rather than successes, and it has no fixed number of trials. When a problem says until, think geometric; when it says in a sample of, decide between the other two by asking whether the items go back.
Section
Part 6
Ranking
Put the steps in the order you would carry them out on a new problem. Some orders are forced by what each step needs.
Put in order
Order the steps for answering a hypergeometric probability question.
Why: The model must be settled before any parameter is written, and X must be defined in words before r can be chosen, because r is whatever group X counts. The values must be listed before the event is translated, so that impossible values never enter the sum. Computing comes after the setup, and interpretation last.
The step most often skipped is the second. Writing X in words is what fixes which group is r, and every later number depends on that choice.
The first step also has a companion question you ask at the same moment: is the sample a tiny fraction of the pool? If so, you may switch to a binomial with p equal to r over r plus b, and the rest of the list changes accordingly.
The fourth step is the one the book's examples teach through Example 4.23. For a small problem the formula gives zero for impossible values anyway, but when you translate at most or more than into a list of terms, knowing the range stops you adding terms that cannot exist.
Pattern
Every problem in this section runs through the same four boxes, with one side question about the binomial.
Figure (svg): A flow diagram of four boxes with arrows: name the group the question counts and write r, b, n; list the values from max(0, n − b) to min(n, r); count C(r, x) C(b, n − x) over C(r + b, n); interpret, with the mean nr/(r + b). A box below asks whether the sample is under 5 percent of the population, in which case a binomial will do.
| step | what you write | Example 4.25 |
|---|---|---|
| name | X in words; r, b, n | men; 6, 5, 4 |
| list | max(0, n − b) to min(n, r) | 0 to 4 |
| count | C(r,x) C(b,n−x) / C(r+b,n) | 15 · 10 / 330 ≈ 0.4545 |
| interpret | a sentence in context; μ = nr/(r+b) | about 2.18 men expected |
The side question: if the sample is under 5 percent of the population, a binomial with p equal to r over r plus b gives almost the same numbers. Otherwise, count.
Three free checks travel with the pattern and catch most errors: the tops of the counts add to the pool and the bottoms to the sample; every value you add lies inside the range; and the answer looks plausible next to the mean, with values near the mean likely and values far from it rare.
Check
One probability, from scratch. Run the bookkeeping check before you trust any of the choices.
Check your understanding
A jar holds 8 red and 12 blue marbles. Five marbles are drawn without replacement. What is the probability that exactly three are red?
Answer: A
Why: Choose three of the eight red marbles in 56 ways and the other two of the twelve blue in 66 ways, then divide by the 15,504 ways to draw any five of the twenty. The tops add to 20 and the bottoms to 5, so the bookkeeping checks.
The binomial choice is close to the right answer, which is exactly what makes it dangerous. Five marbles from twenty is a sampling fraction of a quarter, and the two models happen to agree to one decimal place only by coincidence here.
If you picked the answer with the huge numerator, the check on the bottoms would have caught it: three plus three is six, not the five marbles drawn.
As a final sense test, compare with the mean: five draws with 8 of 20 marbles red gives an expected two red. Three red is one above that, so a probability around a quarter is believable, and 0.79 for a single value is not.
Check
Now the range. Compute both ends from the group sizes.
Check your understanding
A committee of six is chosen from 4 men and 7 women. X is the number of men on the committee. Which values can X take?
Answer: A
Why: Only four men exist, so X cannot exceed 4. There are seven women, enough to fill all six seats, so an all-women committee is possible and X can be 0.
The upper limit comes from the group being counted and the lower limit from the other group. Swapping them is the usual slip, and it produces the range that starts at 2.
A good habit: say each limit as a sentence. At most four, because only four men exist; at least none, because seven women can fill six seats.
Now change the pool to 4 men and 5 women. Six seats and only five women force at least one man onto the committee, and the range becomes 1 to 4. One changed number moves the bottom limit, which is exactly why you compute it every time.
Check
Finally the modelling judgement. The evidence here is the sampling fraction.
Check your understanding
A survey selects 1,000 adults at random, without replacement, from a city of 2 million, and counts how many support a new park. Which statement is correct?
Answer: A
Why: The sampling fraction is 1,000 over 2,000,000, which is 0.0005, far below the five-percent rule of thumb. Each pick changes the chance for the next by a negligible amount, so the binomial with the city's proportion of supporters works well.
Notice that the correct statement names its evidence: the sampling fraction, with a number. A modelling claim with no number behind it earns little credit, however true it is.
If you chose the answer about means, look back at the comparison matrix: the binomial and hypergeometric means agree whenever p is the pool's proportion. The spread is where the two models differ.
And if you chose the answer that rules the binomial out entirely, notice what that would cost: every opinion poll ever taken samples without replacement. The exact model is hypergeometric; the binomial is the accurate and far simpler stand-in whenever the fraction is tiny.
Explain it to yourself
The committee of four has mean 2.18 under both models, but a standard deviation of about 0.83 without replacement and about 1.00 with it. Explaining that split is the best test of whether the section has landed.
Discussion prompt
In two or three sentences, explain why sampling without replacement leaves the mean unchanged but makes the spread smaller. Use the committee of four as your example.
A strong answer separates the two ideas: each pick on its own, which gives the mean, and the picks together, which give the spread. The phrase each pick on its own is the key to the first; pushback or self-correcting is the key to the second.
If your answer only quoted the formulas, add the reason. Could someone who has never seen the formulas predict from your sentences which model has the fatter tails? If yes, you have explained it.
Exit ticket
A committee of five is chosen at random from 7 women and 5 men. You are interested in the number of women.
Discussion prompt
Write the distribution of X, list its values, find the probability of exactly three women and the expected number of women, and say in a sentence whether a binomial would have been acceptable.
For the range: five men are just enough to fill five seats, so zero women is possible, and seven women are more than five, so the top is five. Check your counts: tops 7 and 5 make 12, bottoms 3 and 2 make 5.
A sense check before you hand it in: the mean is about 2.9 women, so three women is the value right next to the mean and should be the most likely single value. A probability of 0.44 fits that; a probability of 0.04 would not.
The final sentence earns the marks. It gives the probability and the mean in terms of women on this committee, and it justifies the model choice with the sampling fraction, a number, not a feeling.
Recap
A hypergeometric experiment draws a sample without replacement from two groups and counts the items from the group the question asks about.
| idea | formula | read it as |
|---|---|---|
| notation | X ~ H(r, b, n) | r in the group of interest, b in the other group, n drawn |
| values | max(0, n − b) ≤ x ≤ min(n, r) | either group can run out |
| probability | C(r,x) C(b,n−x) / C(r+b,n) | favourable samples over all samples |
| mean | μ = nr/(r + b) | sample size times the share of the pool |
| binomial shortcut | n/N < 0.05 | then p = r/(r + b) barely changes |
\[ \binom{r}{x}\binom{b}{n-x} \quad \text{tops add to } r+b, \; \text{bottoms add to } n \]
Every example in the section followed the same path: name the group the question counts, list the values the group sizes allow, count favourable samples over all samples, and say what the number means for the committee, the shipment or the gross.
The picks are dependent, so the binomial formula is wrong for small pools, but each pick on its own has the pool's proportion as its chance, so the mean matches the binomial's and only the spread shrinks. Next, Section 4.6 meets the Poisson distribution, which counts events in an interval rather than items in a sample.
OpenStax Introductory Statistics 2e, §4.5 Hypergeometric Distribution §4.5, pp. 247-250 — the whole section
OpenStax Introductory Business Statistics 2e, §4.1 Hypergeometric Distribution §4.1 — the same distribution in the business-statistics edition
Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.