Inference from a Sample Proportion
Aim
• To be able to draw conclusions about a
population proportion on the basis of
estimates from a sample
Objectives
• Calculate a confidence interval for a
proportion
• Interpret a confidence interval for a
proportion
• Test the hypothesis that the population
proportion is equal to a specific value
• Interpret a P-value from a hypothesis test
on a proportion
This session should take you about 30
minutes to an hour to complete.
Introduction
In this and the following session we will
extend the methods of statistical inference
to the analysis of proportions. We will
calculate a confidence interval for a
proportion and test hypotheses about
values that it may take.
It is common in epidemiology to have
categorical data.
Exercise
What is the term for data which can take one of
two values?
When we analyse data with just two categories we
usually work with proportions (or percentages).
Exercise
The length of stay in days after giving birth was
recorded amongst a group of 200 women
attending a London maternity hospital. The
results are shown opposite.
What proportion of women stayed in hospital for
10 or more days?
Notation
When we want to find out the distribution of
a binary variable, like hypertension in new
mothers, we take a sample of size n from
the population of all new mothers. A
proportion p of all mothers in the sample will
have the characteristic of interest.
To draw conclusions about the population
proportion we then apply the same methods
used to make inferences about sample
means.
The table below shows the notation for population
parameters and sample statistics for proportions.
Remember we use Greek letters for the population
and Roman letters for the sample.
In session SC06 we saw how the
sampling distribution of a proportion is
approximately Normal when the sample is
large.
In that session we also learnt that the SE
of a sample estimate is equal to the
standard deviation divided by √n. Click
here to review this.
Example
For the example of live births in a district hospital and
mothers with hypertension.
The proportion of mothers with hypertension was:
329 / 1310 = 0.251
The standard error is:
SE = √ (0.251 (1 – 0.251) / 1310)
= √ (0.251 x 0.749 / 1310)
SE = 0.012
95% Confidence Interval for a Proportion
If the sample size is large, we can calculate a
confidence interval for a sample proportion using
the usual formula.
For a 95% confidence interval
proportion ± 1.96 x SE (proportion)
If we are working with percentages instead of
proportions, how would this formula change? Think
about it and then press the button below to see the
answer.
Example
In a survey of 335 men on a Health Register in a
small town, 127 said they were current smokers
The estimated proportion of smokers is
127 / 335 = 0.379 (37.9%).
The estimated standard error of the proportion of
smokers is:
A confidence interval for the proportion is given by
0.379 - 1.96 x 0.027 to 0.379 + 1.96 x 0.027
0.326 to 0.432
(32.6% to 43.2%)
Exercise 1
Click on the box (below) which has the
correct interpretation of this confidence
interval.
Exercise 2
A random sample of 100 adults aged 60 or over
living in a small town in the north of England were
interviewed. 22% reported that they suffered from
bronchitis.
What is the standard error of proportion, to
3 decimal places?
Exercise 3
The table opposite shows results from a
survey for filariasis in which 787 men were
examined. The proportion (and 95%
confidence intervals) with hydrocoele by
age group are shown.
We can display these data graphically.
Click the button below to view either the
graph or the table.
The circles show the proportion with hydrocoele,
and the horizontal bars show the 95% confidence
intervals.
In the next session we will look at how to test the
difference in two groups formally.
From the graph do you think there is evidence of
age-related variation in the prevalence of
hydrocoele?
Significance Test for a Single Proportion
If we want to assess whether the population
proportion has a certain value, we use the
procedure for hypothesis testing from SC07.
1. First we should state the Null Hypothesis:
2. Then we state the Alternative Hypothesis:
3. Finally we compute the test statistic. The formula
for a test statistic is the same but with proportions
instead of means.
where p0 is the hypothesised proportion and p the
observed sample proportion.
You will remember from SC07 that the value z
represents the number of standard errors between
the observed and hypothesised proportions.
Example
For a sample of 100 adults aged 60 or over, 22%
reported that they had a medical diagnosis of
bronchitis. The district medical officer was
surprised at this result since she expected this
proportion to be 10%.
We can test whether the observed value is
significantly different from the expected value.
• The null hypothesis is
• The alternative hypothesis is
The P-value for z = 4.0 is 0.00006. This is
interpreted in the following way.
If the proportion of adults with bronchitis in this
population is 10% then the chance of observing a
sample proportion equal or more extreme than
22% is 0.06 in 1000.
So there is strong evidence against the Null
Hypothesis and we conclude that sampling
variation is an unlikely explanation for the
discrepancy between 22% and 10%.
Exercise
In a sample of 150 individuals from a study
of elderly people in the UK, 95 (63.3%) were
female. Does this indicate that there are
more women than men in the elderly
population of the UK?
Click on the statement on the right that
correctly states the null hypothesis for this
question.
Remembering that there are 150 individuals in the
sample, and 63.3% are female, fill in the boxes
below to test whether the proportion of women in
the elderly population of the UK is different from
the proportion of men.
First enter the proportion of women as measured
from the sample, and state what proportion we
expect for the null hypothesis. Give your answers
to 3 decimal places.
Now calculate the standard error of p
assuming the null hypothesis to be true:
The corresponding p-value for z = 3.24, read from 2-
sided tables of the standard Normal distribution, is
0.0012 (see, for example, Altman's book, p. 518 -
520).
So if the percentage of females in the elderly
population of the UK were 50%, then the chance of
observing a sample percentage of 63.3% or more
extreme is 1.2 per 1000 samples.
We conclude that we can reject the Null Hypothesis.
This is an example where we might test whether
the proportion of females among the elderly is 50%
against the alternative hypothesis that it is greater
than 50%.
So,
The test statistic would be the same,
z = 3.24, but the P-value should be read from one-
sided tables (see for example Altman's book: pp.
515-517).
The one-sided P-value is:
P = 0.0006
Small Samples
The methods for making an inference from a
sample population described in this session
are only reliable if the sample is large.
A "rule of thumb" for deciding whether a sample size
is large enough for the distribution of the sample
proportions to be Normal is the following.
Two conditions need to be satisfied by the data. For
indicating the population proportion and n the
sample size, the conditions are:
This happens, for example, when
20 < n < 50 and 0.30 < p < 0.70.
'Exact' tests
When these conditions are not satisfied, we
use the binomial distribution.
The calculations for this test are fairly
complex and will not be covered in this study
unit. It is, however, made available by many
statistical packages.
Summary
The main points of this session will appear
below as you click through the step card
opposite. Click on any of the list entries
below to go back to that card.
Categorical data in Epidemiology
Categorical data are presented as proportions or
percentages.
This session focused on data with just two
categories.
For proportions we use the notation below.
Standard error for a proportion
The standard error for a proportion is given by:
where
p is the sample proportion
and
n is the sample size
If we use percentages instead of proportions we
substitute 100 for 1:
Confidence interval for a proportion
A 95% confidence interval for a proportion is given
by:
proportion ± 1.96 x SE (proportion)
This gives the range of values in which we can be
95% confident the true value lies.
Significance test for a proportion
We can test a specific hypothesis about a
population proportion using the following formula:
and refer z to two-sided tables of the Standardised
Normal distribution to obtain a P-value.
The P-value is the probability of obtaining the
observed, or more extreme, data when the null
hypothesis is true.
Validity of Methods
The methods introduced in this session are
only valid if the following conditions are
satisfied:
πn > 5
(1 –π) n > 5
Exercises
Practical Session 9 (WB09) in the Statistics with
Computing Workbook should now be completed.