Probability and Statistics: To p, or not to p?
Module Leader: Dr James Abdey
4.6 Confidence intervals
A point estimate (such as a sample mean, x̄) is our ‘best guess’ of an unknown population
parameter (such as a population mean, µ) based on sample data. Although:
E(X̄) = µ
meaning that on average the sample mean is equal to the population mean, as it is based on a
sample there is some uncertainty (imprecision) in the accuracy of the estimate. Different
random samples would tend to lead to different observed sample means. Confidence intervals
communicate the level of imprecision by converting a point estimate into an interal estimate.
Formally, an x% confidence interval covers the unknown parameter with x% probability over
repeated samples. The shorter the confidence interval, the more reliable the estimate.
As we shall see, this is achievable by:
• reducing the level of confidence (undesirable)
• increasing the sample size (costly).
If we assume we have either i. known σ, or ii. unknown σ but a large sample size, say
n ≥ 50, then the formulae for the endpoints of a confidence interval for a single mean are:
σ s
i. x̄ ± z × √ and ii. x̄ ± z × √ .
n n
Here x̄ is the sample mean, σ is the population standard deviation, s is the sample standard
deviation, n is the sample size and z is the confidence coefficient, reflecting the confidence
level.
1
Influences on the margin of error
More simply, we can view the confidence interval for a mean as:
best guess ± margin of error
where x̄ is the best guess, and the margin of error is:
σ s
i. z×√ and ii. z×√ .
n n
Therefore, we see that there are three influences on the size of the margin of error (and
hence on the width of the confidence interval). Specifically:
• other things equal, larger sample sizes improve the precision of the point estimate, hence
the confidence interval becomes shorter, so:
as n ↑ ⇒ margin of error ↓ ⇒ width ↓
• other things equal, σ (or s) reflects the amount of variation in the population so a larger
standard deviation means more uncertainty in the representativeness of a random sample,
hence the confidence interval becomes longer, so:
as σ ↑ ⇒ margin of error ↑ ⇒ width ↑
• other things equal, a greater level of confidence equates to a larger confidence coefficient,
hence the confidence interval becomes longer, so:
as confidence level ↑ ⇒ margin of error ↑ ⇒ width ↑ .
Confidence coefficients
For a 95% confidence interval, z = 1.96, leading to:
σ s
i. x̄ ± 1.96 × √ and ii. x̄ ± 1.96 × √ .
n n
Other levels of confidence pose no problem, but require a different confidence coefficient. For
large n, we obtain this coefficient from the standard normal distribution.
• For 90% confidence, use the confidence coefficient z = 1.645.
• For 95% confidence, use the confidence coefficient z = 1.960.
• For 99% confidence, use the confidence coefficient z = 2.576.
2
Example
A company producing designer label jeans carries out a sampling exercise in order to estimate
the average price which all retailers are charging for the jeans.
A random sample of retailers, with an assumed σ = 3.25 gave the following summary statistics:
x̄ = £25.75 and n = 60.
A 95% confidence interval for the mean retailer’s price of the jeans is:
3.25
25.75 ± 1.96 × √ ⇒ (£24.93, £26.57).
60
Note how if the same x̄ was obtained from a random sample of n = 100, then the 95% confidence
interval becomes shorter:
3.25
25.75 ± 1.96 × √ ⇒ (£25.11, £26.39).
100
For the original sample size of n = 60, if instead we had assumed σ = 3.75, then the 95%
confidence interval becomes longer:
3.75
25.75 ± 1.96 × √ ⇒ (£24.80, £26.70).
60
For the original sample size of n = 60 and assumed σ = 3.25, then a 99% confidence interval
becomes longer:
3.25
25.75 ± 2.576 × √ ⇒ (£24.67, £26.83).
60
See how z, σ and n each affect the width of the confidence interval as expected.
