The Sample Proportion
The sample proportion as a random variable, its mean p and standard deviation √(p(1 − p)/n), how its distribution narrows and becomes approximately normal for large samples, and the normal approximation with its limitations.
45 min · Statistical Analysis
Assumes you have read The Binomial Distribution and The Normal Distribution and z-Tables.
The full teaching. Start here.
The short version
- , where counts the sample members with the attribute.
- changes from sample to sample; is a fixed feature of the population.
- and , so .
- Larger samples give less variation: times the sample size halves .
- For large , is approximately normal. Rule of thumb: and .
- Standardise with , then use the empirical rule (, , ) or a table.
- Limitations: is discrete; the approximation is poor for small or near or ; the normal curve extends outside .
A random variable from sampling
Suppose a proportion of a population has some attribute, such as supporting a policy. We usually cannot survey everyone, so we take a random sample of size and use the fraction of the sample with the attribute.
- Sample proportion
, where is the number of members of a random sample of size with the attribute. Because different samples contain different people, is a random variable. Its possible values are .
If the sample is random and selections are independent (sampling with replacement, or from a population much larger than the sample), then , and each probability for is the matching binomial probability: .
CheckpointAnswer before reading on.
In a random sample of size , members have a certain attribute. What is the sample proportion ?
Hint 1Proportion = count ÷ size.
Hint 2X is a count.
Hint 3.
CheckpointAnswer before reading on.
In a random sample of 60 students, 18 walk to school. Find the sample proportion who walk.
Hint 1Count ÷ sample size.
Hint 2.
Hint 30.3.
CheckpointAnswer before reading on.
Why is called a random variable?
Hint 1Take another sample.
Hint 2Does the answer change?
Hint 3Sample to sample.
CheckpointAnswer before reading on.
For samples of size , which values can take?
Hint 1X takes 0 to 5.
Hint 2Divide each by 5.
Hint 3Steps of 0.2.
CheckpointAnswer before reading on.
Samples of size are taken from a population with . Write down the probability distribution of .
Hint 1.
Hint 2.
Hint 3.
Exploring with data and graphs
The best way to see how behaves is to take many samples, collected or simulated, and graph the results with a dot plot or histogram.
Worked example4 marks
Graphing sample proportions from supplied data
In a large school, of students catch a bus. Twenty students each survey a random sample of students and record the proportion who catch a bus. Their results:
| Frequency |
(a) Describe the shape, centre and spread of a dot plot of these values. (b) The mean of the values is and their standard deviation is about . Compare these with the theory.
(a) Shape
Read the frequencies as dot heights.
The dots pile up in the middle at and tail off evenly on both sides: roughly symmetric and single-peaked.
(a) Centre and spread
Where it balances and how far it stretches.
Centred near ; values range from to . Different samples gave different answers: varies between samples.
(b) Theory
Use E and σ for p̂.
and .
(b) Compare
How well does the data match?
The observed mean is very close to . The observed spread is a little under , which is reasonable for only samples; with more samples the two would get closer.
Answer
(a) Roughly symmetric and single-peaked, centred at about , spread from to . (b) Mean ; standard deviation , close to the theoretical .
Is that answer sensible?
The frequencies add to , one value per student.
A simulation does the same thing much faster: a spreadsheet or calculator can generate thousands of samples. Repeating it for different sample sizes shows the pattern in the figure below.
CheckpointAnswer before reading on.
Many samples of size 10, then of size 100, are taken and a histogram of p̂ is drawn for each size. How does the second histogram compare?
Hint 1Spread depends on n.
Hint 2Shape for large n.
Hint 3Same centre.
Exam question
Standard · about 5 min
3 marks
A spreadsheet simulates samples of size from a population with and records for each. (a) Around what value should the values be centred? (b) Predict their standard deviation. (c) Roughly what fraction of the values would you expect between and ?
Hint 1.
Hint 2.
Hint 3.
Where students lose marks on this one
Expecting every simulated p̂ to equal 0.3.
Why it happens: Ignoring sampling variation.
Individual samples vary around p.
Written for this site.
CheckpointAnswer before reading on.
Samples of size are taken from a population with . What shape is the distribution of ?
Hint 1np = 1.
Hint 2p̂ cannot go below 0.
Hint 3Piled up near 0.
Mean and standard deviation of p̂
Since and :
The first result says is centred on the true proportion: on average, sampling neither overestimates nor underestimates. The second says the spread shrinks as grows, like .
CheckpointAnswer before reading on.
In a population, of voters support a policy. Samples of voters are taken. Find .
Hint 1.
Hint 2Sample size does not matter.
Hint 30.35.
CheckpointAnswer before reading on.
For samples of size from a population with , find .
Hint 1.
Hint 2.
Hint 30.0016.
CheckpointAnswer before reading on.
For samples of size from a population with , find the standard deviation of .
Hint 1.
Hint 2.
Hint 3.
CheckpointAnswer before reading on.
The sample size is multiplied by 4. What happens to the standard deviation of p̂?
Hint 1.
Hint 2.
Hint 3Halves.
Exam question
Harder · about 3 min
2 marks
A pollster assumes and wants the standard deviation of to be at most . Find the least sample size.
Hint 1.
Hint 2.
Hint 3.
Where students lose marks on this one
Answering 12.5.
Why it happens: Not squaring 0.02.
Square both sides first.
Written for this site.
Exam question
Standard · about 3 min
2 marks
Given and , show that and .
Hint 1.
Hint 2.
Hint 3.
Where students lose marks on this one
Writing Var(X/n) = Var(X)/n.
Why it happens: Not squaring the constant.
Var(aX) = a² Var(X).
Written for this site.
Common mistake
For n = 100, p = 0.2: Var(p̂) = np(1 − p) = 16.
Why it happens: That is the variance of the count X, not of the proportion.
Dividing by divides the variance by : and . A standard deviation bigger than is impossible for a proportion, which is a quick warning sign.
The normal approximation
For large samples, the distribution of is approximately normal with mean and standard deviation . A common rule of thumb for "large enough" is
This is a guide, not a theorem, and some texts use . When is near or , you need a larger before the skew disappears.
To find a probability, standardise and use the standard normal distribution:
Worked example4 marks
Using the normal approximation
About of a city's households have solar panels. A random sample of households is taken. (a) Check that a normal approximation is reasonable. (b) Estimate . (c) Estimate .
(a) Rule of thumb
Expected successes and failures.
and . Reasonable.
Parameters
Mean and standard deviation of p̂.
and .
(b) Empirical rule
0.25 and 0.35 are one standard deviation from 0.3.
to , so .
(c) Two standard deviations
z = (0.4 − 0.3)/0.05 = 2.
About lies within , leaving in each tail: .
Answer
(a) and , both at least . (b) About . (c) About .
Is that answer sensible?
Probabilities for in the lower tail below should also be about by symmetry, and , which is consistent with the empirical rule.
CheckpointAnswer before reading on.
Which rule of thumb is commonly used to decide whether the normal approximation to p̂ is reasonable?
Hint 1Expected successes.
Hint 2Expected failures.
Hint 3Both at least 5.
CheckpointAnswer before reading on.
For which sample is the normal approximation to least appropriate?
Hint 1Compute np and n(1 − p).
Hint 2Look for one under 5.
Hint 3Extreme p.
CheckpointAnswer before reading on.
A fair coin is tossed times. Using the normal approximation and the empirical rule, estimate .
Hint 1.
Hint 2 and are .
Hint 3About 95%.
CheckpointAnswer before reading on.
Samples of are taken from a population with . Using the normal approximation and , estimate .
Hint 1.
Hint 2.
Hint 3.
CheckpointAnswer before reading on.
Samples of are taken from a population with . Using the normal approximation and the empirical rule, estimate .
Hint 1.
Hint 2.
Hint 3Half of .
CheckpointAnswer before reading on.
For samples of from a population with , find an interval that should contain about of sample proportions.
Hint 1.
Hint 2.
Hint 3.
Exam question
Harder · about 5 min
3 marks
A machine produces defective items. Random samples of items are checked. Using the normal approximation and , estimate the probability that a sample has more than defective.
Hint 1.
Hint 2.
Hint 3.
Where students lose marks on this one
Answering 0.8508.
Why it happens: Giving the wrong tail.
More than means the upper tail.
Written for this site.
Exam question
Harder · about 5 min
3 marks
A company claims of its customers are under . A random sample of customers has under . Is this sample result unusual if the claim is true? Justify.
Hint 1.
Hint 2.
Hint 3.
Where students lose marks on this one
Concluding the claim is proven false.
Why it happens: Overclaiming.
An unusual result casts doubt; it does not prove.
Written for this site.
Try the rule of thumb for yourself. Set and , so : the bars are badly skewed and the normal curve fits poorly. Raise until and the fit improves.
- normal, mean 0.30, sd 0.072
- Mean of p̂
- 0.30
- Standard deviation √(p(1 − p)/n)
- 0.0725
- np and n(1 − p)
- 12.0 and 28.0
- Normal approximation reasonable?
- yes, both are at least 5
Every slider is a normal range input, so the arrow keys move it one step and Home and End jump to the extremes.
Limitations
The normal approximation is a model, and it can fail:
- Discreteness. only takes values in steps of , but the normal curve is continuous. For small this makes a real difference, especially for probabilities like that start exactly at a possible value.
- Skew. For small , or close to or , the true distribution is lopsided, and a symmetric curve fits badly.
- Impossible values. The normal curve gives some probability below and above , where can never be.
- Sampling assumptions. The formulas assume a random sample with independent selections. A biased sample, or a large sample drawn without replacement from a small population, breaks them.
Exam question
Harder · about 5 min
3 marks
A fair coin is tossed times. (a) Find the exact value of . (b) The normal approximation gives about . Explain why the two answers differ so much.
Hint 1.
Hint 2.
Hint 3Small n, discrete.
Where students lose marks on this one
Only computing P(X = 8).
Why it happens: Missing 9 and 10.
"At least" includes all larger values.
Written for this site.
Exam question
Standard · about 4 min
3 marks
Give three limitations of using a normal distribution to approximate the distribution of p̂.
Hint 1Discrete vs continuous.
Hint 2Small n or extreme p.
Hint 3Range of p̂.
Where students lose marks on this one
Saying the approximation never works.
Why it happens: Overstating the limitation.
It is good for large n with p not extreme.
Written for this site.
What lifts this to a Band 6
In a normal approximation question, write three things before any arithmetic: the rule-of-thumb check with numbers (, ), the mean , and with its value. Then draw a quick bell curve and shade the region asked for. The shading catches wrong-tail errors, which are the most common way to lose the final mark.
Why the normal shape appears
is the average of independent Bernoulli variables, each or . A general result, the central limit theorem, says that the average of many independent copies of almost any random variable is approximately normal. The Bernoulli case is the simplest example: each piece looks nothing like a bell, but the average does once is large.
The same law explains a practical fact about polls. Since , the standard deviation of is at most . A sample of about gives at most about , so about of samples land within roughly percentage points of the truth. To halve that margin you need four times as many people, which is why polls rarely go much bigger.