Scope Papers

(collection of statistical papers first

published in Scope)
Edited by J V Freeman

Page 1
Page 2
 Contents
The visual display of quantitative information ........................................................................5
Describing and summarising data ......................................................................................... 13
The Normal distribution .......................................................................................................21
Hypothesis testing and Estimation ........................................................................................ 27
Randomisation in Clinical Investigations .............................................................................. 35
Basic tests for continuous data .............................................................................................. 43
Basic tests for continuous data: Mann-Whitney U and Wilcoxon signed rank sum tests ........49
The analysis of categorical data ............................................................................................ 55
The analysis of categorical data: Fisher’s Exact test .............................................................. 61
Use of Statistical Tables .......................................................................................................65
One-way Analysis of Variance ............................................................................................. 73
Sample Size Calculations for Clinical Trials: parallel, two-group clinical trial with a Normally
distributed outcome variable. ................................................................................................ 79
Diagnostic Tests ................................................................................................................... 87
Correlation Coefficient .........................................................................................................93
Simple Linear Regression ................................................................................................... 101
Risk .................................................................................................................................... 107
References .......................................................................................................................... 114

Page 3
Page 4
The visual display of quantitative information
Dr Jenny V. Freeman, Steven A. Julious

Introduction

A knowledge of the appropriate methods for displaying data and results is a valuable
skill for researchers and scientists. Appropriate figures are useful as they can be read
quickly, and are particularly helpful when presenting information to an audience. In
addition, plotting data is an extremely useful first stage to any analysis, as this could
show extreme observations (outliers) together with any interesting patterns. In this
tutorial, good practice guidelines for presentation of data in figures are given. In
addition the basic types of data are described, together with some simple figures
appropriate for their display.

Good practice recommendations for construction of figures

Box 1 outlines some basic recommendations for the construction and use of figures
to display data. A fundamental principle, for both figures and tables is that they
should maximise the amount of information presented for the minimum amount of ink
used(Tufte 1983). Good figures have the following four features in common: clarity of
message, simplicity of design, clarity of words and integrity of intentions and
action(Bigwood and Spore 2003). A figure should have a title explaining what is
displayed and axes should be clearly labelled; if it is not immediately obvious how
many individuals the figure is based upon, this should also be stated. Gridlines
should be kept to a minimum as they act as a distraction and can interrupt the flow of
information. When using figures for presentation purposes care must be taken to
ensure that they are not misleading; an excellent exposition of the way in which
figures can be used to mislead can be found in Huff(Huff 1991).

Box 1 : Guidelines for good practice when constructing figures

1. The amount of information should be maximised for the minimum amount of
ink
2. Figures should have a title explaining what is being displayed
3. Axes should be clearly labelled
4. Gridlines should be kept to a minimum
5. Avoid 3-D charts as these can be difficult to read
6. The number of observations should be included

Types of data

In order to appropriately display data, it is first important to understand the different

types of data there are as this will determine the best method of displaying them.
Briefly, data are either categorical or quantitative. Data are described as
categorical when they can be categorised into distinct groups, such as ethnic group
or disease severity. Categorical data can be divided into either nominal or ordinal.
Nominal data have no natural ordering and examples include eye colour, marital
status and area of residence. Binary data is a special subcategory of nominal data,

Page 5
where there are only two possible values, for example (male/female, yes/no,
treated/not treated). Ordinal data occurs when there can be said to be a natural
ordering of the data values, such as better/same/worse, grades of breast cancer,
social class or quality scores.

Quantitative variables can be either discrete or continuous. Discrete data are also
known as count data and occur when the data can only take whole numbers, such as
the number of visits to a GP in a year or the number of children in a family.
Continuous data are data that can measured and they can take any value on the
scale on which they are measured; they are limited only by the scale of measurement
and examples include height, weight, blood pressure, area or volume.

Basic charts for categorical data

Categorical data may be displayed using either a pie chart or a bar chart. Figure 1
shows a pie chart of the distribution of marital status by sex for UK adults at the 2001
census. Each segment of the pie chart represents the proportion of the UK
population who are in that category. It is clear from this figure that differences
between the sexes exist with respect to marital status; nearly half of all men have
never married, whilst this proportion was smaller for women. Interestingly the
proportion of women who were widowed was about three times that for men. Figure 2
displays the same data in a bar chart. The different marital status categories are
displayed along the horizontal axis whilst on the vertical axis is percentage. Each bar
represents the percentage of the total population in that category. For example,
examining Figure 2, it can be seen that the percentage of men who are married is
about 48%, whilst the percentage of women is closer to 40%. Generally pie charts
are to be avoided as they can be difficult to interpret particularly when the number of
categories becomes greater than 5. In addition, unless the percentages in the
individual categories are displayed (as here) it can be much more difficult to estimate
them from a pie chart than from a bar chart. The relative proportions falling in the
different categories is much clearer in Figure 2 than in Figure 1. For both chart types
it is important to include the number of observations on which it is based, particularly
when comparing more than one chart. And finally, neither of these charts should be
displayed as 3-D as these are especially difficult to read and interpret.

Page 6
Figure 1: Pie Chart of marital status for UK,
http://www.statistics.gov.uk/STATBASE/Expodata/Spreadsheets/D7680.xls

Men (n=28,579,900) Women (n=30,209,300)

Widow ed (2.9)
Divorced (5.6) Widow ed (10.3)

Separated (1.7)
Divorced (7.1)
Remarried (5.9)
Separated (2.3)
Never married (40.9)
Remarried (5.4)

Never married (48.0)

Married (35.9)

Married (34.0)

Figure 2: Data in Figure 1 displayed as a Bar Chart

60

Men
50
Women

40
Percentage

30

20

10

0
gle st) rri e
d ed ed
we
d
Sin (f i r rat orc ido
d ma pa i v
rri e Re Se D W
Ma

Page 7
 Basic charts for quantitative data

There are several charts that can be used for quantitative data. Dot plots are one of
the simplest ways of displaying all the data. Figure 3 shows dot plots of the heights
for a random sample of 100 couples. Each dot represents the value for an individual
and is plotted along a vertical axis, which in this case, represents height in metres.
Data for several groups can be plotted alongside each other for comparison; for
example, data for the 100 randomly sampled couples are plotted separately by sex in
Figure 3 and the differences in height between men and women can be clearly seen.

Figure 3: Dot plot of heights of random sample of 100 couples

200

190

180

170

160
Height (cm)

150

140

Men (n=100) Women (n=100)

A common method for displaying continuous data is a histogram. In order to

construct a histogram the data range is divided into several non-overlapping equally
sized categories and the number of observations falling into each category counted.
The categories are then displayed on the horizontal axis and the frequencies
displayed on the vertical axis, as in Figure 4. Occasionally the percentages in each
category are displayed on the y-axis rather than the frequencies and it is important
that if this is done, the total number of observations that the percentages are based
upon must be included in the chart. The choice of number of categories is important
as too few categories and much important information is lost, too many and any
patterns are obscured by too much detail. Usually between 5 and 15 categories will
be enough to gain an idea of the distribution of the data.

Page 8
 Figure 4: Histograms of height for random sample of 100 couples, by sex

Men (n=100) Women (n=100)

30 30

20 20

10 10
Percentage

Percentage
0 0
14

15

15

15

16

16

16

17

17

17

18

18

18

19

19

14

15

15

15

16

16

16

17

17

17

18

18

18

19

19
8.

1.

5.

8.

1.

4.

7.

1.

4.

7.

0.

3.

7.

0.

3.

8.

1.

5.

8.

1.

4.

7.

1.

4.

7.

0.

3.

7.

0.

3.
6

4
Height (cm) Height (cm)

A useful feature of a histogram is that it is possible to assess the distributional form of

the data; in particular whether the data are approximately Normal, or are skewed.
The histogram of Normally distributed data will have a classic „bell‟ shape, with a
peak in the middle and symmetrical tails, such as that for men in Figure 4a. The
Normal distribution (sometimes known as the Gaussian distribution) is one of the
fundamental distributions of statistics, and its properties, which underpin many
statistical methods, will be discussed in a later tutorial. Skewed data are data which
are not symmetrical, negatively skewed data have a long left-hand tail at lower
values, with a peak at higher values, whilst conversely positively skewed data have a
peak at lower values and a long tail of higher values.

Another extremely useful method of plotting continuous data is a box-and-whisker

or box plot (Figure 5). Box plots can be particularly useful for comparing the
distribution of the data across several groups. The box contains the middle 50% of
the data, with lowest 25% of the data lying below it and the highest 25% of the data
lying above it. In fact the upper and lower edges represent a particular quantity called
the interquartile range. The horizontal line in the middle of the box represents the
median value, the value such that half of the observations lie below this value and
half lie above it. The whiskers extend to the largest and smallest values excluding the
outlying values. The outlying values are those values more than 1.5 box lengths from
the upper or lower edges, and are represented as the dots outside the whiskers.
Figure 5 shows box plots of the heights of the men and women. As with the dot plots,
the gender differences in height are immediately obvious from this plot and this
illustrates the main advantage of the box plot over histograms when looking at
multiple groups. Differences in the distributions of data between groups are much
easier to spot with box plots than with histograms.

Page 9
 Figure 5: Boxplot of height

200

190

180

170

160
Height (cm)

150

140
men (n=100) w omen (n=100)

The association between two continuous variables can be examined visually by

constructing a scatterplot. The values of one variable are plotted on the horizontal
axis (known as the X-axis) and the values of another are plotted on the vertical axis
(Y-axis). If it is known (or suspected) that the value of one variable (independent)
influences the value of the other variable (dependent), it is usual to plot the
independent variable on the horizontal axis and the dependent variable on the
vertical axis. Although it is not always obvious, it is often clear which variables to
place on the X- and Y-axes. Experimentally the X-axis would be something that the
experimenter controls while the Y-axis would be the response to the X-axis.
Variables that tend to go on the X-axis are variables such age, time (hours, months,
years etc) or temperature. This will be discussed again in greater detail in a later
tutorial examining the techniques of regression and correlation. Figure 6 shows the
scatter plot of women‟s height against their partner‟s height and each dot represents
the two height values for a couple. Note, here variables could have been placed
interchangeably on the X-or Y-axes.

Summary

This tutorial has given basic good practice guidelines for producing figures and
highlighted some of the simple figures available for displaying data. The list of figures
described is not exhaustive and the work presented here will be revisited and
extended in subsequent tutorials. And finally, it is worth considering that often the
simplest plots convey the strongest message.

Page 10
Figure 6: Scatter plot of heights for 100 couples, women on horizontal axis,
men on vertical axis

200

190

180

170

160

150

140

130
130 140 150 160 170 180 190 200

Women, height (cm)

Page 11
Page 12
Describing and summarising data
Jenny V Freeman and Steven A Julious

Introduction

A recent tutorial in SCOPE gave some good practice recommendations for the visual
display of quantitative data and highlighted some simple figures available for
displaying data(Freeman and Julious 2005c). This tutorial is concerned with ways of
describing and summarising data. In addition, the presentation of numbers and use
of tables will also be covered and good practice guidelines for communicating
quantitative information will be outlined.

Describing categorical data

An initial step when describing categorical data is to count the number of

observations in each category and express them as percentages of the total sample
size. For example, Table 1 shows the marital status of the UK population taken from
the 2002 census by sex(Anon 2005). The data are categorised in two ways, by
marital status and gender, enabling the distribution of marital status to be compared
between the two sexes; Table 1 is an example of a contingency table with 6 rows
(representing marital status) and 2 columns (representing gender) and marital status
is said to have been cross-tabulated with study group. When presenting data in this
way (as percentages), it is important to include the denominator for each group (total
sample size), as giving percentages alone can be misleading if the groups contained
very different numbers(Altman and Bland 1996).

Table 1: Marital status for UK population, 2001 census

Men Women
(n=28,579,900) (n=30,209,300)
Never married 48.0 40.9
Married 35.9 34.0
Divorced 5.6 7.1
Remarried 5.9 5.4
Separated 1.7 2.3
Widowed 2.9 10.3

Describing quantitative data

As it can be difficult to make sense of a large amount of quantitative data, an initial

approach, in addition to examining the data visually, is to calculate summary
measures, to describe the location (a measure of the „middle value‟) and the spread
(a measure of the dispersion of the values) for each variable. These are of great
interest, particularly if a comparison between groups is to be made or the results of
the study are to be generalised to a larger group.

Measures of location

There are several measures of location, as summarised in Box 1. The simplest is the
mode. This is simply the most common observation and is the highest bar of the

Page 13
histogram. Looking at the histogram for height of a random sample of 100 men, the
modal height around 171.9cm as this is the height category with the highest bar on
the histogram (Figure 1a). However, the mode is rarely used since its value depends
upon the accuracy of the measurement. If, for example, the number of height bands
on the histogram were increased from 14 to 19, the mode would change to 176cm
(Figure 1b). In addition, it can be difficult to determine if there is more than one
distinct peak – for example two peaks would indicate bi-modal data. The most
practical use for a mode is when the variable is either ordinal or discrete (with a finite
number of categories) where one category dominates.

Box 1: Measures of location

Mode Most common observation

Median Middle observation, when the data are arranged in order of increasing
value. If there is an even number of observations the median is
calculated as the average of the middle two observations
n

x
= Sum of all observations
i
Mean = x = i 1
Number of observations
n
where x is the sample mean, xi is the ith observation, n is the sample
n
size and the notation represents the addition or summing up of all
i 1
the observations from the first (i = 1) to the last (n).

For example, consider the ages (in years) of five individuals: 42,32, 41, 45 and 38.

The most common observation is: 42, 32, 41, 45 or 38. Unfortunately, multiple modes
exist in this example, so there is no unique mode.

The five ages in ascending order are: 32, 38, 41, 42, and 45. The median is the
middle or 3rd value of the ranked or ordered ages i.e. 41 years.

The mean is: 82 + 72 + 81 + 85 + 58 = 198 divided by the number of observations, 5,

i.e. 39.6 years.

Page 14
 Figure 1. Histograms of height for a random sample of 100 men, (a) with 14
bins (categories/bars) and (b) with 19 bins. Note the change in the value of the
highest bar
(a) (b)
20 20

Frequency
Frequency

10 10

0 0

15
16 0
16 0
16 .0
16 0
16 0
17 0
17 0
17 0
17 .0
17 0
18 0
18 0
18 0
18 0
18 0
19 .0
19 0
19 0
15

16

16

16

16

17

17

18

18

18

18

19

19
17

8.
0.
2
4.
6.
8.
0.
2.
4
6.
8.
0.
2.
4.
6.
8
0.
2.
4.
8.

1.

3.

6.

9.

1.

4.

0.

2.

5.

8.

0.

3.
7.

0
4

6
4

Height (cm) Height (cm)

Two other more useful measures are the median and the mean. The median is the
middle observation, when the data are arranged in order of increasing value. It is the
value that divides the data into two equal halves. If there is an even number of
observations then the median is calculated as the average of the two middle
observations. For example, if there are 11 observations the median is simply the 6 th
observation, but if there are 10 observations the median is the (5 th + 6th
observation)/2. The median is not sensitive to the behaviour of outlying data, thus if
the smallest value was even smaller, or the largest value even bigger it would have
no impact on the value of the median. The median height for the 100 men is
176.3cm.

Probably the most commonly used measure of the central value of a set of data is
the mean. It is calculated as the sum of all observations divided by the total number
of observations. Each observation makes a contribution to the mean value and thus it
is sensitive to the behaviour of outlying data; as the largest value increases this
causes the mean value to increase and conversely, as the value of the smallest
observation becomes smaller the value of the mean decreases. The mean height for
the 100 men is 176.5cm.

Both the mean and median can be useful, but they can give very different
impressions when distribution of the data is skewed, because of the relative
contributions (or lack of, in the case of the median) of the extreme values. Skewed
data are data that are not symmetrical and this is best illustrated by examining the
histograms for such data, as in Figures 2a and 2b. Data, such as that in Figure 2a,
which have a long right-hand tail of higher values, but where the majority of
observations are clustered at lower values are called right skewed, or positively
skewed (conversely data where the observations are clustered at higher values but
with a long left-hand tail of lower values such as that in Figure 2b are called left
skewed or negatively skewed).

Page 15
 Figure 2: Distribution of data displaying positive (a) and negative (b) skew.
Data are taken from a randomised controlled trial examining the cost
effectiveness of community leg ulcer clinics (n=233)(Morrell et al. 1998).
(a) (b)
50 50

40 40
Frequency

Frequency
30 30

20 20

10 10

0 0
10

10

12

12

13

14
11

14
45

52

59

66

73

80

87

94

30

35

40

45

50

55

60

65

70

75

80

85

90

95
0.

7.

1.

8.

5.

9.
4.

2.
.9

.8

.7

.6

.4

.3

.2

.1

.0

.0

.0

.0

.0

.0

.0

.0

.0

.0

.0

.0

.0

.0
9

1
7

2 Weight (kg) Age (years)

There are no firm rules about which to use, but when the distribution is not skew it is
usual to use the mean; it is worth noting that if the data are symmetrically distributed
the mean and median will be similar, as can be seen from the mean and median of
the height data described earlier (176.3cm and 176.5cm respectively). However, if
data are skew then it is better to use the median, as this is not influenced by the
extreme values and may not be as misleading as the mean; an extreme example
would be the median of the sample 1, 2, 3, 4 and 100,000, which would be 3,
whereas the mean is 20,002. One example of where medians have been used in
preference to means is in reporting salaries. Due to a relatively few outlying high-
income earners the vast majority of workers were earning much less than the mean
wage, thus nowadays, medians are produced and quoted(Bird 2004).

Measures of Spread

In addition to finding measures to describe the location of a dataset, it is also

necessary to be able to describe its spread. Just as with the measures of location,
there are both simple and more complex possibilities (as summarised in Box 2). The
simplest is the range of the data, from the smallest to the largest observation. The
range of height for the random sample of 100 men is 159 to 193cm (or 35 cm as a
single number). The advantage of the range is that it is easily calculated, but its
drawback is that it is vulnerable to outliers, extremely large and extremely small
observations. A more useful measure is to take the median value as discussed
above and further divide the two data halves into halves again. These values are
called the quartiles and the difference between the bottom (or 25% percentile) and
top quartile (or 75th percentile) is referred to as the inter-quartile range (IQR). This is
the observation below which the bottom 25% of the data lie and the observation
above which the top 25% lie: the middle 50% of the data lie between these limits.
Unlike the range it is not as sensitive to the extreme values. The IQR for the height of
the random sample of 100 men is 172.5 to 181 cm (or 8.5cm as a single number).
Strictly speaking the range and IQR are single numbers but frequently the two

Page 16
values, minimum and maximum, or the 25% and 75% percentiles respectively, are all
reported as this can be more informative.

Box 2: Measures of spread

Range Minimum observation to the maximum observation

Interquartile Observation below which the bottom 25% of data lie and the
range observation above which the top 25% of data lie. If the value
falls between two observations, e.g. if 25th centile falls
between 5th and 6th observations then the value is calculated
as the average of the two observation (this is the same
principle as for the median

n
( xi x) 2
i 1
Standard deviation = SD =
n 1

where x is the sample mean, xi is the ith observation, n is the

n
sample size and the notation represents the addition or
i 1
summing up of all the squared deviations from the sample
mean from the first (i = 1) to the last (nth) observation.

For example, consider the ages (in years) of the five individuals abovel: 32, 38, 41,
42 and 45.

The range of the data is from 32 to 45 years or 13 years.

The five ages in ascending order are: 32, 38, 41, 42 and 45. The bottom 25% of data,
falls somewhere between the 1st and 2nd ordered observations, i.e. 32 and 38, so we
can take the average of these two observations 32 + 38 = 70/2 = 35 years. The top
25% of data, falls somewhere between the 4th and 5th ordered observations, i.e. 42
and 45. So the 75th percentile is the average of the two observations 42 + 45 = 87/2 =
43.5. Hence the interquartile range is 35.0 to 43.5 years or 8.5 years.

The standard deviation is calculated by first working out the squared deviation of
each observation from the sample mean of 39.6 years i.e.
(32 – 39.6)2 + (38 – 39.6)2 + (41 – 39.6)2 + (42 – 39.6)2 + (45 – 39.6)2 = 92.7 years2.
This result is divided by the number in the sample minus one (i.e. 5 - 1= 4) i.e. 97.2/4
= 24.3 years2. Finally, we take the square root of this number to give us a standard
deviation of 4.93 years.

The most common measure of the spread of the data is the standard deviation (see
Box 2) and it is used in conjunction with the mean. It provides a summary of the
differences of each observation from the mean value. The standard deviation (SD)
has units on the same scale as the original measurement (e.g. cm if height is being
measured in cm). For the sample of 100 men, the SD for height is 6.6cm and
provides a measure of average deviation for an observation from the sample mean.

Page 17
As with the measures of location, when deciding which measure of spread to present
it is necessary to know whether the data are skewed or not; this will also have a
bearing on how the data will be analysed subsequently, as will be seen in the
following chapter. When the distribution is not skewed it is usual to use the standard
deviation. However, if data are skewed then it is better to use the range or inter-
quartile range.

Presentation of numbers

As with charts, there are a few basic rules of good presentation, both within the text
of a document or presentation, and within tables, as outlined in Box 3.

Box 3: Guidelines for good practice when presenting numbers:

1. The amount of information should be maximised for the minimum amount of

ink.
2. Numerical precision should be consistent throughout a paper or presentation,
as far as possible.
3. Avoid spurious accuracy. Numbers should be rounded to two effective digits.
4. Quantitative data should be summarised using either the mean and SD (for
symmetrically distributed data) or the median and IQR or range (for skewed
data). The number of observations on which these summary measures are
based should be included.
5. Categorical data should be summarised as frequencies and percentages. As
with quantitative data, the number of observations should be included.
6. Tables should have a title explaining what is being displayed and columns and
rows should be clearly labelled.
7. Gridlines in tables should be kept to a minimum.
8. Rows and columns should be ordered by size.

A fundamental principle is that the amount of information should be maximised for

the minimum amount of ink(Tufte 1983). For summarising numerical data, the mean
and standard deviation should be used, or if the data have a skewed distribution the
median and range or inter-quartile range should be used. However, for all of these
calculated quantities it is important to state the total number of observations on which
they are based. When summarising categorical data, both frequencies and
percentages can be used, but if percentages are reported it is important that the
denominator (i.e. total number of observations) is given. Numerical precision should
be consistent throughout and summary statistics such as means and standard
deviations should not have more than one extra decimal place (or significant figure)
compared to the raw data. Spurious precision should be avoided, although when
certain measures are to be used for further calculations greater precision may
sometimes be appropriate(Altman & Bland 1996).

Tables, including the column and row headings, should be clearly labelled and a brief
summary of the contents of a table should always be given in words, either as part of
the title or in the main body of the text. Gridlines can be used to separate labels and
summary measures from the main body of the data in a table. However, their use
should be kept to a minimum, particularly vertical gridlines, as they can interrupt eye

Page 18
movements, and thus the flow of information. Elsewhere white space can be used to
separate data, for example, different variables from each other. The information in
tables is easier to comprehend if the columns (rather than the rows) contain like
information, such as means and standard deviations, as it is easier to scan down a
column than across a row(Ehrenberg 2000). Where there is no natural ordering of the
rows (or indeed columns), such as marital status in Table 1, they should be ordered
by size as this helps the reader to scan for patterns and exceptions in the
data(Ehrenberg 2000).

Table or Chart?

Some basic charts for displaying data were described in the previous tutorial. Plotting
data is a useful first stage to any analysis and will show extreme observations
together with any discernible patterns. Charts are useful as they can be read quickly,
and are particularly helpful when presenting information to an audience such as in a
seminar or conference presentation. Although there are no hard and fast rules about
when to use a chart and when to use a table, when presenting the results in a report
or a paper it is often best to use tables so that the reader can scrutinise the numbers
directly. Tables can be useful for displaying information about many variables at
once, whilst charts can be useful for showing multiple observations on groups or
individuals.

Summary

The essence of any attempt to present data and results, either in a presentation or on
paper is to communicate with an audience and it is hoped that by following the basic
rules outlined here that task will be made simpler. This tutorial has covered some
basic measures for describing and summarising data. It has also outlined some good
practice guidelines for communicating quantitative information. The next tutorial will
examine the Normal distribution and the Central Limit Theorem.

Page 19
Page 20
The Normal distribution
Jenny V Freeman, Steven A Julious

Introduction

The first two tutorials in this series have focussed on displaying data and simple
methods for describing and summarising data. There has been little discussion of
statistical theory. In this note we will start to explore some of the basic concepts
underlying much statistical methodology. We will describe the basic theory
underlying the Normal distribution and the link between empirical frequency
distributions (the observed distribution of data in a sample) and theoretical probability
distributions (the theoretical distribution of data in a population). In addition we will
introduce the idea of a confidence interval.

Theoretical probability distributions

Since it is rarely possible to collect information on an entire population, the aim of

many statistical analyses is to use information from a sample to draw conclusions (or
„make inferences‟) about the population of interest. These inferences are facilitated
by making assumptions about the underlying distribution of the measurement of
interest in the population as a whole, by applying an appropriate theoretical model to
describe how the measurement behaves in the population1. In the context of this note
the population is a theoretical concept used for describing an entire group and one
way of describing the distribution of a measurement in a population is by use of a
suitable theoretical probability distribution. Probability distributions can be used to
calculate the probability of different values occurring and they exist for both
continuous and categorical measurements.

In addition to the Normal distribution (described later in this note), there are many
other theoretical distributions, including the Chi-squared, Binomial and the Poisson
distributions (these will be discussed in later tutorials). Each of these theoretical
distributions is described by a particular mathematical expression (formally referred
to as a model) and for each model there exist summary measures, known as
parameters which completely describe that particular distribution. In practice,
parameters are usually estimated by quantities calculated from the sample, and
these are known as statistics i.e. a statistic is a quantity calculated from a sample in
order to estimate an unknown parameter in a population. For example, the Normal
distribution is completely characterised by the population mean (μ) and population
standard deviation (σ) and these are estimated by the sample mean ( x ) and sample
standard deviation (s) respectively.

1
Note that prior to any analysis it is usual to make assumptions about the underlying
distribution of the measurement being studied. These assumptions can then be investigated
through various plots and figures for the observed data – for example a histogram for
continuous data. These investigations are referred to as diagnostics and will be discussed
throughout subsequent notes.

Page 21
The Normal distribution

The Normal, or Gaussian distribution (named in honour of the German

mathematician C.F.Gauss, 1777-1855) is the most important theoretical probability
distribution in statistics. At this point it is important to stress that in this context the
word „Normal‟ is a statistical term and is not used in the dictionary or clinical sense of
conforming to what would be expected. Thus, in order to distinguish between the two,
statistical and dictionary „normal‟, it is conventional to use a capital letter when
referring to the Normal distribution.

Box 1: Properties of the Normal distribution

1. It is bell shaped and has a single peak (unimodal)

2. Symmetrical about the mean
3. Uniquely defined by two parameters, the mean (μ) and standard deviation (σ)
4. The mean, median and mode all coincide
5. The probability that a Normally distributed random variable, x, with mean, μ,
and standard deviation, σ, lies between the limits (μ – 1.96σ) and (μ + 1.96σ)
is 0.95
i.e. 95% of the data for a Normally distributed random variable will lie between
the limits (μ – 1.96σ) and (μ + 1.96σ)*
6. The probability that a Normally distributed random variable, x, with mean, μ,
and standard deviation, σ, lies between the limits (μ – 2.44σ) and (μ + 2.44σ)
is 0.99
7. Any position on the horizontal axis of Figure 1 can be expressed as a number
of standard deviations away from the mean value

*This fact is used for calculating the 95% confidence interval for Normally distributed
data

The basic properties of the Normal distribution are outlined in box 1. The distribution
curve of data which are Normally distributed has a characteristic shape; it is bell-
shaped, and symmetrical about a single peak (figure 1). For any given value of the
mean, populations with a small standard deviation have a distribution clustered close
to the mean (μ), whilst those with a large standard deviation have a distribution that is
widely spread along the measurement axis and the peak is more flattened.

As mentioned earlier the Normal distribution is described completely by two

parameters, the mean (μ) and the standard deviation (σ). This means that for any
Normally distributed variable, once the mean and variance (σ2) are known (or
estimated), it is possible to calculate the probability distribution for that population.

Page 22
Figure 1: The Normal Distribution

An important feature of a Normal distribution is that 95% of the data fall within 1.96
standard deviations of the mean – the unshaded area in the middle of the curve on
figure 1. A summary measure for a sample often quoted is the two values associated
with the mean +/- 1.96 x standard deviation ( x +/-1.96s). These two values are
termed the Normal range and represent the range within which 95% of the data are
expected to lie. Note 68.7% of data lie within 1 standard deviation of the mean whilst
virtually all of the data (99.7%) will lie within 3 standard deviations (95.5% will lie
within 2). The Normal distribution is important as it underpins much of the
subsequent statistical theory outlined both in this and later tutorials, such as the
calculation of confidence intervals and linear modelling techniques.

The Central Limit Theorem (or the law of large numbers)

The Central Limit Theorem states that given any series of independent, identically
distributed random variables, their means will tend to a Normal distribution as the
number of variables increases. Put another way, the distribution of sample means
drawn from a population will be Normally distributed whatever the distribution of the
actual data in the population as long as the samples are large enough.

In order to illustrate this, consider the random numbers 0 to 9. The distribution of

these numbers in a random numbers table would be uniform. That is to say that each
number has an equal probability of being selected and the shape of the theoretical
distribution is represented by a rectangle. According to the Central Limit Theorem, if
you were to select repeated random samples of the same size from this distribution,
and then calculate the means of these different samples, the distribution of these
sample means would be approximately Normal and this approximation would
improve as the size of each sample increased. Figure 2a represents the distribution
of the sample means for 500 samples of size 5. Even with such a small sample size
the approximation to the Normal is remarkable, whilst repeating the experiment with
samples of size 50, improves the fit to the Normal distribution (Figure 2b). The other
noteworthy feature of these two figures is that as the size of the samples increases
(from 5 to 50), the spread of the means is decreased.

Page 23
Figure 2: Distribution of means from 500 samples

(a) Samples of size 5, mean=4.64, (b) Samples of size 50, mean=4.50,

sd=1.29 sd=0.41

Each mean estimated from a sample is an unbiased estimate of the true population
mean and by repeating the sampling many times we can obtain a sample of plausible
values for the true population mean. Using the Central Limit Theorem we can infer
that 95% of sample means will lie within 1.96 standard deviations of the population
mean. As we do not usually know the true population mean the more important
inference is that with the sample mean we are 95% confident that the population
mean will fall within 1.96 standard deviations of the sample mean. In reality, as we
usually only take a single sample, we can use the Central Limit Theorem to construct
an interval within which we are reasonably confident the true population mean will lie.
This range of plausible values is known as the confidence interval and the formula
for the confidence interval for the mean is given in Box 2. Technically speaking, the
95% confidence interval is the range of values within which the true population mean
will lie 95% of the time if a study was repeated many times. Crudely speaking, the
confidence interval gives a range of plausible values for the true population mean.
We will discuss confidence intervals further in subsequent notes in context with
hypothesis tests and P-values.

Box 2: Formula for the confidence interval for a mean

s s
x 1.96 to x 1.96
n n

s = sample standard deviation and n = number of individuals in the sample

In order to calculate the confidence interval we need to be able to estimate the

standard deviation of the sample mean. It is defined as the sample standard
deviation, s, divided by the square root of the number of individuals in the sample,
s n and is usually referred to as the standard error. In order to avoid confusion, it
is worth remembering that using the standard deviation (of all individuals in the

Page 24
sample) you can make inferences about the spread of the measurement within the
population for individuals whilst using the standard error you can make inference
about the spread of the means: the standard deviation is for describing (the spread
of data) whilst the standard error is for estimating (how precisely the mean has been
pinpointed).

Summary

In this tutorial we have outlined the basic properties of the Normal distribution and
discussed the Central Limit Theorem and outlined its importance to statistical theory.
The Normal distribution is fundamental to many of the tests of statistical significance
outlined in subsequent tutorials, whilst the principles of the Central Limit Theorem
enable us to calculate confidence intervals and make inference about the population
from which the sample is taken.

Page 25
Page 26
Hypothesis testing and Estimation
Jenny V Freeman, Steven A Julious

Introduction

In the previous tutorial we outlined the basic properties of the Normal distribution and
discussed the Central Limit Theorem(Freeman and Julious 2005b). The Normal
distribution is fundamental to many of the tests of statistical significance covered in
subsequent tutorials. As a result of the principles of the Central Limit Theorem the
Normal distribution enables us to calculate confidence intervals and make inference
about the population from which the sample is taken. In this tutorial we explain the
basic principles of hypothesis testing (using P-values) and estimation (using
confidence intervals). By the end of the tutorial you will know of the processes
involved and have an awareness of what a P-value is and what it is not, and what is
meant by the phrase „statistical significance‟.

Statistical Analysis

It is rarely possible to obtain information on an entire population and usually data or

information are collected on a sample of individuals from the population of interest.
Therefore one of the main aims of statistical analysis is to use this information from
the sample to draw conclusions („make inferences‟) about the population of interest.

Consider the hypothetical example of a study designed to examine the effectiveness

of two treatments for migraine. In the study patients were randomly allocated to two
groups corresponding to either treatment A or treatment B. It may be that the primary
objective of the trial is to investigate whether there is a difference between the two
groups with respect to migraine outcome; in this case we could carry out a
significance test and calculate a P-value (hypothesis testing). Alternatively it may be
that the primary objective is to quantify the difference between treatments together
with a corresponding range of plausible values for the difference; in this case we
would calculate the difference in migraine response for the two treatments and the
associated confidence interval for this difference (estimation).

Hypothesis Testing (using P-values)

Figure 1 describes the steps in the process of hypothesis testing. At the outset it is
important to have a clear research question and know what the outcome variable to
be compared is. Once the research question has been stated, the null and alternative
hypotheses can be formulated. The null hypothesis (H0) usually assumes that there is
no difference in the outcome of interest between the study groups. The study or
alternative hypothesis (H1) usually states that there is a difference between the study
groups.

In lay terms the null hypothesis is what we are investigating whilst the alternative is
what we often wish to show. For example when comparing a new migraine therapy
against control we are investigating whether there is no difference between
treatments. We wish to prove that this null hypothesis is false and demonstrate that
there is a difference at a given level of significance.

Page 27
Figure 1: Hypothesis testing: the main steps

In general, the direction of the difference (for example: that treatment A is better than
treatment B) is not specified, and this is known as a two-sided (or two-tailed) test. By
specifying no direction we investigate both the possibility that A is better than B and
the possibility that B is better than A. If a direction is specified this is referred to as a
one-sided test (one-tailed) and we would be evaluating only whether A is better then
B as the possibility of B being better than A is of no interest. It is rare to do a one-
sided test as they have no power to detect a difference if it is in the opposite direction
to the one being evaluated. We will not dwell further on the difference between two-
sided and one-sided tests other than to state that the convention for one-sided tests
is to use a level of significance of 2.5% - half that for a two-sided test. Usually in
studies it is two-sided tests that are done.

A common misunderstanding about the null and alternative hypotheses, is that when
carrying out a statistical test, it is the alternative hypothesis (that there is a difference)
that is being tested. This is not the case – what is being examined is the null
hypothesis, that there is no difference between the study groups; we conduct a
hypothesis test in order to establish how likely (in terms of probability) it is that we
have obtained the results that we have obtained, if there truly is no difference in the
population.

For the migraine trial, the research question of interest is:

„For patients with chronic migraines which treatment for migraine is the most
effective?‟

There may be several outcomes for this study, such as the frequency of migraine
attacks, the duration of individual attacks or the total duration of attacks. Assuming
we are interested in reducing the frequency of attacks, then the null hypothesis, H o,
for this research question is:

„There is no difference in the frequency of attacks between treatment A and

treatment B groups’

Page 28
and the alternative hypothesis, H1, is:

„There is a difference in the frequency of attacks between the two treatment

groups‟.

Having set the null and alternative hypotheses the next stage is to carry out a
significance test. This is done by first calculating a test statistic using the study data.
This test statistic is then compared to a theoretical value under the null hypothesis in
order to obtain a P-value. The final and most crucial stage of hypothesis testing is to
make a decision, based upon the P-value. In order to do this it is necessary to
understand first what a P-value is and what it is not, and then understand how to use
it to make a decision about whether to reject or not reject the null hypothesis.

So what does a P-value mean? A P-value is the probability of obtaining the study
results (or results more extreme) if the null hypothesis is true. Common
misinterpretations of the P-value are that it is either the probability of the data having
arisen by chance or the probability that the observed effect is not a real one. The
distinction between these incorrect definitions and the true definition is the absence
of the phrase when the null hypothesis is true. The omission of „when the null
hypothesis is true‟ leads to the incorrect belief that it is possible to evaluate the
probability of the observed effect being a real one. The observed effect in the sample
is genuine, but what is true in the population is not known. All that can be known with
a P-value is, if there truly is no difference in the population, how likely is the result
obtained (from the sample). Thus a small P-value indicates that difference we have
obtained is unlikely if there genuinely was no difference in the population – it gives
the probability of obtaining the study results (or results more extreme) (difference
between the two study samples) if there actually is no difference in the population.

In practice, what happens in a trial is that the null hypothesis that two treatments are
the same is stated i.e. A=B or A-B=0. The trial is then conducted and a particular
difference, d, is observed where A-B=d. Due to pure randomness even if the two
treatments are the same you would seldom observe A-B=0. Now if d is small (say a
1% difference in the frequency of attacks) then the probability of seeing this
difference under the null hypothesis is very high say P=0.995. If a larger difference is
observed then the probability of seeing this difference by chance is reduced, say
d=0.05 then the P-value could be P=0.562. As the difference increases therefore so
the P-value falls such that a d=0.20 may equate to a P=0.021. This relationship is
illustrated in Figure 2: as d increases then the P-value (under the null hypothesis)
falls.

Page 29
Figure 2. Illustration of the relationship between the observed difference and
the P-value under the null hypothesis

It is important to remember that a P-value is a probability and its value can vary
between 0 and 1. A „small‟ P-value, say close to zero, indicates that the results
obtained are unlikely when the null hypothesis is true and the null hypothesis is
rejected. Alternatively, if the P-value is „large‟, then the results obtained are likely
when the null hypothesis is true and the null hypothesis is not rejected. But how
small is small? Conventionally the cut-off value or two-sided significance level for
declaring that a particular result is statistically significant is set at 0.05 (or 5%).
Thus if the P-value is less than this value the null hypothesis (of no difference) is
rejected and the result is said to be statistically significant at the 5% or 0.05 level
(Box 1). For the example above, if the P-value associated with the mean difference in
the number of attacks was 0.01, as this is less than the cut-off value of 0.05 we
would say that there was a statistically significant difference in the number of attacks
between the two groups at the 5% level.

Box 1: Statistical Significance

We say that our results are statistically significant if the P-value is less than
the significance level (α), usually set at 5%

P < 0.05 P≥0.05

Result is Statistically signific nt No statistically
significant
Decide That t re is sufficient That there is insufficient
evidence to reject the null evidence to reject the null
hypothesis and accept the hypothesis
alternative hypothesis

We cannot say that the null

hypothesis is true, only that there
is not enough evidence to reject it

Page 30
The choice of 5% is somewhat arbitrary and though it is commonly used as a
standard level for statistical significance its use is not universal. Even where it is, one
study that is statistically significant at the 5% level is not usually enough to change
practice; replication is required. For example to get a license for a new drug usually
two statistically significant studies are required at the 5% level which equates to a
single study at the 0.00125 significance level. It is for this reason that larger „super‟
studies are conducted to get significance levels that would change practice i.e. a lot
less than 5%.

Where the setting of a level of statistical significance at 5% comes from is not really
known. Much of what we refer to as statistical inference is based on the work of R.A.
Fisher (1890-1962) who first used 5% as a level of statistical significance acceptable
to reject the null hypothesis. One theory is that 5% was used because Fisher
published some statistical tables with different levels of statistical significance and
5% was the middle column. An exercise we do with students in order to demonstrate
empirically that 5% is a reasonable level for statistical significance is to toss a coin
and tell the students whether we‟ve observed a head or a tails. We keep saying
heads. After around 6 tosses we ask the students when they stopped believing we
were telling the truth. Usually about half would say after 4 tosses and half after 5. The
probability of getting 4 heads in a row is 0.063 and the probability of getting five
heads in a row is 0.031; hence 5% is a figure about which most people would
intuitively start to disbelieve an hypothesis!

Although the decision to reject or not reject the null hypothesis may seem clear-cut, it
is possible that a mistake may be made, as can be seen from the shaded cells of Box
2. For example a 5% significance level means that we would only expect to see the
observed difference (or one greater) 5% of the time under the null hypothesis.
Alternatively we can rephrase this to state that even if the two treatments are the
same 5% of the time we will conclude that they are not and we will make a Type I
error. Therefore, whatever is decided, this decision may correctly reflect what is true
in the population: the null hypothesis is rejected, when it is fact false or the null
hypothesis is not rejected, when in fact it is true. Alternatively, it may not reflect what
is true in the population: the null hypothesis may be rejected, when in fact it is true
which would lead us to a false positive and making a Type I error, (α); or the null
hypothesis may not be rejected, when in fact it is false. This would lead to a false
negative, and making a Type II error, (β). Acceptable levels of the Type I and Type
II error rates are set before the study is conducted. As mentioned above the usual
level for declaring a result to be statistically significant is set at a two sided level of
0.05 prior to an analysis i.e. the type I error rate (α) is set at 0.05 or 5%. In doing this
we are stating that the maximum acceptable probability of rejecting the null when it is
in fact true (committing a type 1 error, α error rate) is 0.05. The P-value that is then
obtained from our analysis of the data gives us the probability of committing a Type I
error (making a false positive error).

Page 31
Box 2: Making a decision

The null hypothesis is actually:

The P value. This
Decide to: False True is the probability of
Reject the null hypothesis Correct Type 1 Error (α) concluding that
Not reject the null Type 2 Error Correct there is a
hypothesis (β) difference, when in
fact there is no
This represents a well powered study – difference, i.e. the
one that is able to detect a difference when probability of
there truly is a difference making a false
positive mistake

The probability that a study will be able to detect a difference, of a given size, if one
truly exists is called the Power of the study and is the probability of rejecting the null
hypothesis when it is actually false (probability of making a Type II error, β). It is
usually expressed in percentages, so for a study which has 90% power, there is a
probability of 0.9 of being able to detect a difference, of a given size, if there
genuinely is a difference in the population. An underpowered study is one which
lacks the ability, i.e. has very low power, to detect a difference when there truly is a
difference. The concepts of power and Type I and II errors will be dealt with further in
a later tutorial on sample size, as these are important components of sample size
calculation.

Estimation (using confidence intervals)

Statistical significance does not necessarily mean the result obtained is

clinically significant or of any practical importance. A P-value will only indicate
how likely the results obtained are when the null hypothesis is true. It can only be
used to decide whether the results are statistically significant or not, it does not give
any information about the likely size of the clinical difference. Much more information,
such as whether the result is likely to be of clinical importance can be gained by
calculating a confidence interval. Although in the previous tutorial we talked about
the 95% confidence interval for the mean, it is possible to calculate a confidence
interval for any estimated quantity (from the sample data), such as the mean,
median, proportion, or even a difference. It is a measure of the precision (accuracy)
with which the quantity of interest is estimated (in the case of the migraine trial, the
quantity of interest is the mean difference in the number of migraine attacks).

Technically, the 95% confidence interval is the range of values within which the true
population quantity would fall 95% of the time if the study were to be repeated many
times. Crudely speaking, the confidence interval gives a range of plausible values for
the quantity estimated; although not strictly correct it is usually interpreted as the
range of values within which there is 95% certainty that the true value in the
population lies. For the migraine example, let us assume that the quantity estimated,
the mean difference in the number of attacks between the groups, was 3 attacks per
month and the 95% confidence interval for this difference was 1.2 to 4.8 attacks per
month. Thus, whilst the best available estimate of the mean difference was 3 attacks
Page 32
per month, it could be as low as 1.2 or as high as 4.8 attacks per month, with 95%
certainty. As the confidence interval excludes 0 we can infer from the observed trial
that it is unlikely that there is no difference between treatments. In fact as we have
calculated a 95% confidence interval we can deduce that the statistical significance is
less than 5%. The actual P-value associated with this difference was 0.01 and given
that it is less than 5% we can conclude that the difference is statistically significant at
the 5% level.

As confidence intervals are so informative and from them we can infer statistical
significance as well as quantify plausible values for the population effect there is a
growing consensus that only confidence intervals should be reported for studies.
However, it is unlikely that P-values will ever be eliminated as a way to quantify
differences.

Statistical and Clinical Significance

So far in this tutorial we have dealt with hypothesis testing and estimation. However,
in addition to statistical significance, it is useful to consider the concept of clinical
significance. Whilst a result may be statistically significant, it may not be clinically
significant (relevant/important) and conversely an estimated difference that is
clinically important may not be statistically significant. For example consider a large
study comparing two treatments for high blood pressure; the results suggest that
there is a statistically significant difference (P=0.023) in the amount by which blood
pressure is lowered. This P-value relates to a difference of 3mmHg between the two
treatments. Whilst the difference is statistically significant, it could be argued that a
difference of 3mmHg is not clinically important. This is supported but the 95%
confidence interval of 2.3 to 3.7mmHg. Hence, although there is a statistically
significant difference this difference may not be sufficiently large enough to convince
anyone that there is a truly important clinical difference.

This is not simply a trivial point. Often in presentations or papers P-values alone are
quoted and inferences about differences between groups are made based on this
one statistic. Statistically significant P-values may be masking differences that have
little clinical importance. Conversely it may be possible to have a P-value greater
than the magic 5% but for there to be a genuine difference between groups: absence
of evidence does not equate to evidence of absence.

Summary

In this tutorial we have outlined the basic principles of hypothesis testing (using P-
values) and estimation (using confidence intervals). In subsequent tutorials we will
be applying this knowledge when performing statistical significance testing in order to
make decisions about the results of analyses.

Page 33
Page 34
Randomisation in Clinical Investigations
Steven A. Julious and Jenny V. Freeman

Introduction

In previous notes we have outlined methods for describing and summarising data,
and the principles of hypothesis testing and estimation(Freeman and Julious
2005a;Freeman and Julious 2006b). In this note we will describe the basic concepts
of randomisation in investigations. We will begin by describing the background,
followed by describing the rationale for randomisation and then finally will move on to
some of the more advanced topics of randomisation pertinent to imaging
investigations.

Background

Allocation at random has been a central tenet of clinical trial design since the first
reported modern clinical trial was conducted to investigate the effect of streptomycin
and bed rest compared to bed rest alone in the treatment of tuberculosis(Bradford
Hill 1990;Julious and Zariffa 2002;Medical Research Council 1948;Yoshioka 1998).
Randomisation is important as it ensures that the regimen groups being investigated
are objectively the same for any demographic or prognostic factors. Randomisation
achieves this by ensuring that each subject has a known chance of receiving a given
treatment in an allocation that can not be predicted(Altman and Bland 1999). This
lack of predictability is important as an investigator should remain masked to the
order of the treatments in order to reduce the potential for bias; only finding out what
regimen a patient is to be assigned after recruiting a patient into the trial(Day and
Altman 2000).

Note the concept of randomisation originally came from clinical trials hence the
reference to treatments. As we will describe in this note though, randomisation is an
important consideration for all types of clinical investigation. The problem of not
allocating at random is evidenced by the following example(Julious and Mullee
1994). An historical trial was undertaken to compare the success of a new treatment
[percutaneous nephrolithomy] with an existing treatment [open surgery] in the
removal of kidney stones. Examining table 1a it appears that the new treatment is
superior with an 83% success rate compared to only 78% on the old treatment.
However, when we break the table down into small (table 1b) and large stones (table
1c) in each table the direction of the effect first observed is reversed. The old
treatment is superior for both sizes of stones. The only reason why the old treatment
seemed inferior to start with was that treatment is confounded with stone. This
reversal effect is known as Simpson's Paradox(Simpson 1951;Williams 1949)).

Page 35
Table 1. A comparison of the success rates of percutaneous nephrolithotomy
(New) compared to open surgery (Old) in the removal of Kidney stones a)
overall b) for stones <2cm and c) stones ≥2cm
a) Overall
Success
Yes No Total
Treatment New 289 (83%) 61 (17%) 350
Old 273 (78%) 77 (22%) 350
Total 562 138 700

b) Stones <2cm
Success
Yes No Total
Treatment New 234 (83%) 36 (17%) 270
Old 81 (93%) 6 (7%) 87
Total 315 42 357

c) Stones ≥2cm
Success
Yes No Total
Treatment New 55 (69%) 25 (31%) 80
Old 192 (73%) 71 (27%) 263
Total 247 96 343

Confounding is a statistical term for when there is a strong relationship between a

third factor and both the outcome and comparison of interest. In the table people
who had percutaneous nephrolithomy were also more likely to have small stones and
the smaller the stone the better the prognosis. Hence, we would say treatment is
confounded with stone size.

Obviously a bias of the magnitude observed with instances of Simpson's Paradox is

rare but randomisation protects the investigator from confounding with known and
unknown prognostic factors. Therefore, wherever possible, subjects should be
assigned to investigations at random. If there are known factors that could effect the
outcome, such as centre, age, sex, or baseline risk then the study should be stratified
to allow for these and a block size (see below) should be set that provides balance
within each strata (see below). If there is to be a constraint in the randomisation,
such as unequal allocation then this should be allowed for in the block size and
appropriate adjustment made to the sample size. Block size and strata are described
below.

Mechanics of Randomisation

Parallel Group Trials

A parallel group trial is one in which there are to be at least two arms to be
investigated and subjects are to be randomised to each of these arms. It is beyond
the scope of this note to describe in detail how to undertake a randomisation
however we will give some general hints and tips (used in this context arm is a
generic term to describe the groupings in trials. Subjects may be assigned to two

Page 36
different arms where these arms could be: treatments; assessors or imaging
protocols).
When randomising subjects to the different arms in the trial an important
consideration is to maintain balance for the interventions to which subjects are being
randomised. This is particularly important in small studies where by chance there
can easily be an imbalance in the number of subjects on the respective arms. One
tool to ensure that groups are balanced is to do introduce “blocks” into the
randomisation. Basically a block is a sample size after which there is balance in the
randomisation. It is best to illustrate this through a simple worked example.

Consider the case of two groups. We wish to randomly allocate individuals to either
group A or group B. In this example we could toss a coin and record either heads (H)
or tails (T), so that we can then use the order to allocate individuals to groups (i.e. if
heads then group A, if tails then group B. If we set the block size to be 4 we need to
ensure that after every four tosses there are two heads and two tails. Thus:

Block 1: T T (H H)
Block 2: T T (H H)
Block 3: T H T (H)
Block 4: T H H (T)

The terms in brackets are not from tosses but entries we were forced to enter to
ensure balance. For example in Block 1 the first two tosses were tails. We thus
made the next 2 heads so that after “4 tosses” we had a balance. Notice after “16
tosses” by blocking we have 8 heads and 8 tails.

Another important consideration is stratification. Stratification is similar to blocking

but here as well as ensuring balance after a requisite block size we also ensure
balance by strata. These strata are usually clinical important sub groups such as sex
and age.

Again it is best to illustrate by example. Suppose we are doing the same coin tossing
to create a randomisation list. For this randomisation we wish to ensure balance for
a two level stratification factor. Operationally this would be the same as doing the
coin tossing exercise twice: once for each stratum.

Stratum 1
Block 1: T T (H H)
Block 2: T T (H H)

Stratum 2
Block 1: T H T (H)
Block 2: T H H (T)

Now after “16 tosses” we have balance both in terms of heads and tails and also for
heads and tails by strata.

A final consideration, as discussed earlier in the note, is the withholding of the

randomisation until the actual allocation of subjects. Even for completely open
studies it is preferable to mask the randomisation so that investigators only find out
what regimen a patient is to be assigned after the patient has been recruited. In

Page 37
practice this could be done by putting the randomisation in envelopes which are
opened only after a subject has been enrolled.

Cross-over Trials

The distinction between parallel group designs and crossover designs is that in
parallel group designs subjects are assigned at random to receive only one
investigation, and as a result of the randomisation the groups are the same in all
respects other than the investigation made. However, with a cross-over trial all
subjects receive all the investigations but it is the order that subjects receive the
investigations which is randomised. The big assumption here is that prior to starting
each investigation all subjects return to baseline and that the order in which subjects
have their investigation does not affect their response to the investigation.

Two Period

Two period cross over trials are the easiest to explain. In the simplest case, for a two
arm investigation (comparing A with B say) subjects will be randomised to either A
followed by B (AB) or to B followed by A (BA). AB and BA are called sequences and
represent the order in which subjects receive the investigations. In practice, subjects
are randomly assigned to the either the sequence AB or the sequence BA, and to
ensure balance blocking can still be used.

Note that even for retrospective investigations, randomisation should be considered.

For example, in a study to investigate the agreement between two image analysts
the analysts could have the images assessed randomly with the analysts reading the
images in random order much like a AB/BA design.

Multi-period

All investigations are made on all subjects

Imaging comparisons can be complicated as there are often a finite number of

subjects on whom a number of investigations are made, such as:

A fMRI investigation where subjects will receive a number of challenges.

A comparison of different imaging protocols within the same subject.
An assessment of new technology such as a comparison of 2D, 3D and 2D
and 3D combined SPECT.
A comparison of several readers on the same subjects to look at agreement.
A comparison of different therapies or different doses of the same therapy
within a subject.

It is quite easy therefore for 4 or 5 investigations to be made on the same subject. If

4 investigations are made on the same subject, that would result in 24 different ways
of assigning subjects to these four investigations and hence 24 sequences. This is
all very good but what if we have only 12 subjects in the trial?

Actually, for multi period investigations we do not necessarily need to use all possible
sequences but can form special sequences to be randomised called Williams
Squares(Williams 1949)

Page 38
It is again best to illustrate through example. In order to investigate an even number
of investigations we can build a Williams square from the following sequence:

0, 1, t, 2, t-1, 3, t-2…etc

where t is the number of interventions minus 1. If we were to conduct 4 investigations

then t=3 and our sequences would include 0, 1, 2, 3. We build the sequences by
forming the first row from the result above. We then form the second by adding 1 to
this first row, but where the number is 3 the new number becomes 0 (we are adding
in base 3). The calculation is simpler than the explanation

0 1 3 2
1 2 0 3
2 3 1 0
3 0 2 1

This is known as a Latin Square: each investigation appears in every row and
column. The columns here would reflect different imaging sessions. A Williams
Square is special form of Latin Square such that as well as being balanced for rows
and columns each investigation is preceded by each other investigation at least once
e.g. 1 is preceded by 0, 2, and 3. Here we are saying that as well the order of
investigations being important the effect of preceding investigations is too. Hence we
ensure balance for the immediately preceding investigation. This is known as first
order balance.

If we were conducting a trial where we are to undertake four different investigations

on 12 subjects we would randomise the 4 sequences above so each sequences
appears 3 times.

For an odd number of investigations we need to build 2 Latin squares with starting
sequences

0, 1, t, 2, t-1, 3, t-2… etc

and

…t-2, 3, t-1, t, 1, 0.

With 5 investigations, t=4 and we would therefore have

0 1 4 2 3
1 2 0 3 4
2 3 1 4 0
3 4 2 0 1
4 0 3 1 2

Page 39
And

3 2 4 1 0
4 3 0 2 1
0 4 2 3 2
1 0 3 2 3
2 1 3 3 4

Not all investigations are made on all subjects

In imaging investigations there are logistical, practical and safety considerations to be

taken into account. For example we may wish to investigate 4 different imaging
protocols but these must all be done in one day for each subject and for practical
reasons we can only schedule 3 scans in a day. Similar we may wish to look at 4
different protocols but for safety reasons we may only be able to do 3 scans in the 24
hours we have each subject. Although we can still construct Latin Squares we need
to construct a special type of these known as a Balanced Incomplete Block. Again
we will illustrate by example.

If we could have 3 sessions for each subject but we have 4 investigations, then
taking the sequences derived previously and removing the first column

0 1 3 2
1 2 0 3
2 3 1 0
3 0 2 1

and final column

0 1 3 2
1 2 0 3
2 3 1 0
3 0 2 1

would give us 8 sequences as follows

1 3 2
2 0 3
3 1 0
0 2 1

0 1 3
1 2 0
2 3 1
3 0 2

We would hence have balance both for rows and columns as well as for first order
balance within 8 sequences.

Page 40
For an odd number of sequences we use a similar procedure. Using our previous
example of have 5 investigations and assuming that we can only do 3 sessions then
we could delete the last 2 columns off the first 5 sequences and the first 2 columns
off then next i.e.

0 1 4 2 3
1 2 0 3 4
2 3 1 4 0
3 4 2 0 1
4 0 3 1 2

3 2 4 1 0
4 3 0 2 1
0 4 2 3 2
1 0 3 2 3
2 1 3 3 4

Discussion

This note introduced the basic concepts of randomisation including the importance of
stratification and blocking. We then described issues pertinent to imaging
investigations where we may wish to perform multiple investigations on each subject
or the special case where the number of investigations is greater than the number of
sessions.

Page 41
Page 42
Basic tests for continuous data
Jenny V Freeman, Steven A Julious

As it is rarely possible to study an entire population, data are usually collected from a
sample of individuals in order to make inferences, or draw conclusions, about the
population of interest. This can be done through a process known as hypothesis
testing, the basic principles of which have been outlined in a previous
tutorial(Freeman & Julious 2006b). At the outset it is important to have a clear
research question and know what the outcome variable to be compared is. Once the
research question has been stated, the null and alternative hypotheses can be
formulated. The null hypothesis (H0) assumes that there is no difference in the
outcome of interest between the study groups. The study or alternative hypothesis
(H1) states that there is a difference between the study groups. Next the appropriate
statistical test must be selected and conducted to obtain a P-value. This P-value will
then be used to make a decision about whether the results are statistically significant
and whether the null hypothesis can be rejected.

This tutorial will provide a concrete example of how the process of setting and testing
a hypothesis is implemented in practice. It will focus on some elementary methods for
analysing continuous data: the paired and unpaired t-tests and their non-parametric
equivalents. Continuous data are data that can be measured and can take any value
on the scale on which they are measured; examples include height, weight, blood
pressure and area coverage.

Choosing the statistical method

What type of statistical analysis depends on the answer to five key questions (Box 1)
and given answers to these, an appropriate approach to the statistical analysis of the
data collected can be decided upon. The type of statistical analysis depends
fundamentally on what the main purpose of the study is. In particular, what is the
main question to be answered? The data type for the outcome variable will also
govern how it is to be analysed, as an analysis appropriate to continuous data would
be completely inappropriate for binary data. In addition to what type of data the
outcome variable is, its distribution is also important, as is the summary measure to
be used. Highly skewed data require a different analysis compared to data which are
Normally distributed.

Box 1: Five key questions to ask:

1. What are the aims and objectives?

2. What is the hypothesis to be tested?
3. What type of data is the outcome data?
4. How is the outcome data distributed?
5. What is the summary measure for the outcome data?

Before beginning any analysis it is important to examine the data, using the
techniques described in the first two tutorials in this series(Freeman & Julious
2005a;Freeman & Julious 2005c); adequate description of the data should precede
and complement the formal statistical analysis. For most studies and for RCTs in

Page 43
particular, it is good practice to produce a table or tables that describe the initial or
baseline characteristics of the sample.

Comparison of two independent groups

Independent samples t-test

The independent samples t-test is used to test for a difference in the mean value of a
continuous variable between two independent groups. For example, as part of a
randomised clinical trial of two treatments for venous leg ulcers one of the main
questions of interest was whether there was a difference in the number of ulcer free
weeks between the control and the clinic groups(Morrell, Walters, Dixon, Collins,
Brereton, Peters, & Brooker 1998). As the number of ulcer free weeks is continuous
data and there are two independent groups, assuming the data are Normally
distributed in each of the two groups, then the most appropriate summary measure
for the data is the sample mean and the best comparative summary measure is the
difference in the mean number of ulcer free weeks between the two groups. When
conducting any statistical analysis it is important to check that the assumptions which
underpin the chosen method are valid. The assumptions underlying the two-sample t-
test are outlined in Box 2. The assumption of Normality can be checked by plotting
two histograms, one for each sample; these do not need to be perfect, just roughly
symmetrical. The two standard deviations should also be calculated and as a rule of
thumb, one should be no more than twice the other.

Box 2: The assumptions underlying the use of the independent samples t-test:

1. The groups are independent.

2. The variables of interest are continuous.
3. The data in both groups have similar standard deviations.
4. The data is Normally distributed in both groups.

The test statistic for the independent samples t-test, t, is calculated as follows:

x1 x2
t
s 2p s 2p
n1 n2

where x1 and x2 are the means of the two groups and n1 and n2 are the numbers in
n1 1 s12 _ n2 1 s22
the two groups and s 2p = and is an estimate of the pooled
n1 n2 2
variance. Once this test statistics has been calculated it can be compared to values
for the t distribution on n1 n2 2 degrees of freedom. These can either be found in
books of statistical tables, or using the TDIST function in EXCEL. This function
requires 3 arguments, X = the t statistic obtained above, deg_freedom = n1 n2 2
and tails, where 1 indicates a one sided test and 2 indicates a two sided test. It is
recommended that you always use 2 here to ensure that you test is two sided. Note
that t can take negative as well as positive values and as the number of degrees of
freedom gets larger the t distribution approaches the Normal distribution.

Page 44
For the leg ulcer data, there were 120 patients in the clinic group and their mean
number of ulcer free weeks for was 20.1. There were 113 patients in the control
group and they had a mean number of ulcer free weeks was 14.2. The pooled
estimate of variance, s 2p , was 5.57. Putting these into the formula above gives a t
statistic of 2.49 on 231 degrees of freedom and this in turn gives a P-value of 0.014.
The final and most crucial stage of hypothesis testing is to make a decision, based
upon this P value. A P value is the probability of obtaining the study results (or results
more extreme) if the null hypothesis is true. It tells you how likely is the result
obtained (from the study data), if there truly is no difference in the population. A
„small‟ P value, say close to zero, indicates that the results obtained are unlikely
when the null hypothesis is true and the null hypothesis is rejected. Alternatively, if
the P value is „large‟, then the results obtained are likely when the null hypothesis is
true and the null hypothesis is not rejected. Conventionally the cut-off value or
significance level for declaring that a particular result is statistically significant is set at
0.05 (or 5%). Thus if the P value is less than this value the null hypothesis (of no
difference) is rejected and the result is said to be statistically significant at the 5% or
0.05 level. For the example above, of the difference in the number of ulcer free
weeks, the P value is 0.014. As this is less than the cut-off value of 0.05 there is said
to be a statistically significant difference in the number of ulcer-free weeks between
the two groups at the 5% level.

Mann-Whitney U test

There are several possible approaches when at least one of the requirements for the
t-test is not met. The data may be transformed (e.g. the logarithm transformation can
be useful particularly when the variances are not equal) or a non-parametric method
can be used. Non-parametric or distribution free methods do not involve distributional
assumptions i.e. making assumptions about the manner in which the data are
distributed (for example that the data are Normally distributed). An important point to
note is that it is the test that is parametric or non-parametric, not the data.

When the assumptions underlying the t-test are not met, then the non-parametric
equivalent, the Mann-Whitney U test, may be used. Whilst the independent samples
t-test is specifically a test of the null hypothesis that the groups have the same mean
value, the Mann-Whitney U test is a more general test of the null hypothesis that the
distribution of the outcome variable in the two groups is the same; it is possible for
the outcome data in the two groups to have similar measures of central tendency or
location, such as mean and medians, but different distributions.

The Mann-Whitney U test requires all the observations to be ranked as if they were
from a single sample. From this the statistic U is calculated; it is the number of all
possible pairs of observations comprising one from each sample for which the value
in the first group precedes a value in the second group. This test statistic is then used
to obtain a P value. The principle is best illustrated with a simple example. Consider
the following two samples of size 6 X=(0,6,5,1,1,6) and Y=(9,4,7,8,3,5). Rank these
in order as if they were from the same sample:

0 1 1 3 4 5 5 6 6 7 8 9

Page 45
Having ranked the samples, choose one sample, for example X and count the
number of observations from Y that are below each of the X observations. The first X
value is 0 and there are 0 Y observations below this. The next X value is 1 and there
are 0 Y values below this. Where there are ties, these can be disregarded and do not
contribute to the total. Continue for each member of X and when this has been done
the numbers of preceding Ys can be added to obtain the U statistic,
Ux=0+0+0+2+3+3=8. This procedure is repeated for the second sample and then the
smaller of the two numbers is used obtain a P-value. As with the t statistic above this
value is compared to tabulated critical values under the null hypothesis to obtain a P-
value. For the data above, the P-value obtained from the Mann-Whitney U test is
0.12. As this is greater than 0.05 the result is not statistically significant, and there
that there is insufficient evidence to reject the null that the two groups differ in terms
of their location.

Two groups of paired observations

When there is more than one group of observations it is vital to distinguish the case
where the data are paired from that where the groups are independent. Paired data
may arise when the same individuals are studied more than once, usually in different
circumstances, or when individuals are paired as in a case-control study. As part of
leg ulcer trial the researchers were interested in assessing whether there was a
change in health related quality of life (HRQoL) between baseline and 3 months for
those individuals with a healed leg ulcer (irrespective of study group). HRQoL at
baseline and 3 months are both continuous variables and the data are paired as
measurements are made on the same individuals at baseline and 3 months;
therefore, interest is in the mean of the differences not the difference between the
two means.

Paired t-test

If we assume that the paired differences are Normally distributed, then the best
comparative summary measure is the mean of the paired difference in HRQoL
between baseline and 3 months. Given the null hypothesis (H0) that there is no
difference (or change) in mean HRQoL at baseline and 3 months follow-up in
patients whose leg ulcer had healed by 3 months, the most appropriate test is the
paired t-test. There were 36 patients with a healed leg ulcer at 3 months.

d
The test statistic for the paired t test is again t and is calculated as t where d
se d
is the mean of the paired differences and se(d ) is the standard error of d and is
sd (d )
estimated as and n is the number of paired differences. As with the unpaired
n
case this t statistic can then be compared to values for the t distribution on n-1
degrees of freedom. The mean change (3 months – baseline) in HRQoL for the 36
patients with healed ulcers was -7.33 with an SD of 16.5 and using these in the
formulae above gives a t value of -2.661, which in turn gives a P-value of 0.012. As
this is less than the nominal level usually set for statistical significance of 0.05 (or
5%) we can conclude that there is a statistically significant difference in HRQoL
between baseline and 3 months. It is worth noting that as the mean change is
negative HRQoL actually declined for these patients over the 3 months!

Page 46
Box 3: The assumptions underlying the use of the paired t-test

1. The differences are plausibly Normally distributed (it is not essential

for the original observations to be Normally distributed).
2. The differences are independent of each other.

Wilcoxon signed rank sum test

The assumptions underlying the use of the paired t-test are outlined in Box 3. If these
are not met a non-parametric alternative, the Wilcoxon signed rank sum test, can be
used. This test is based upon the ranks of the paired differences and test the null
hypothesis that the median of these differences is 0 i.e that there is no tendency for
the outcome in one group (or under one condition) to be higher or lower than in the
other group (or condition). For the leg ulcer trial the null hypothesis would state that
there is no tendency for HRQol life at baseline to be better or worse than HRQoL at 3
months. An explanation of how to carry out a Wilcoxon signed rank sum test in
practice can be found in Swinscow and Campbell(Swinscow and Campbell 2002).
For the leg ulcer data the P-value is 0.012, very similar to that for the paired t-test
above. As it is less than 0.05 we would conclude that there is sufficient evidence to
reject the null and conclude that the median difference is not equal to 0.

Summary

Outlined above are some simple methods for comparing two groups of continuous
data. However, it is important to bear in mind that statistical significance does not
necessarily mean the result obtained is clinically significant or of any practical
importance. A P value will only indicate how likely the results obtained are when the
null hypothesis is true. Much more information, such as whether the result is likely to
be of clinical importance can be gained by calculating a confidence interval, as this a
range of plausible values for the estimated quantity. In the next tutorial we will extend
the methods outlined above to cover more than two groups.

Page 47
Page 48
Basic tests for continuous data: Mann-Whitney U
and Wilcoxon signed rank sum tests
Jenny V Freeman, Michael J Campbell

The most recent tutorial examined how the process of setting and testing a
hypothesis could be implemented in practice(Freeman and Julious 2006a). It
focussed on some elementary methods for analysing continuous data: the paired and
unpaired t-tests. However, these tests make particular assumptions about the
distribution of the data. Most importantly that the standard deviations are similar (for
the independent groups t-test) and that the data to be analysed are approximately
Normally distributed (both tests).

This tutorial will discuss some alternative methods that can be used when these
assumptions are violated. They are part of a group of statistical tests known as non-
parametric or distribution-free tests; distribution-free tests do not involve making any
assumptions about how the data are distributed (for example that the data are
Normally distributed). An important point to note is that it is the test that is parametric
or non-parametric, not the data.

Mann-Whitney U test

When the assumptions underlying the independent samples t-test are not met, then
the non-parametric equivalent, the Mann-Whitney U test, may be used. Whilst the
independent samples t-test is specifically a test of the null hypothesis that the groups
have the same mean value, the Mann-Whitney U test is not a test for a difference in
medians, as is commonly thought. It is a more general test of the null hypothesis that
the distribution of the outcome variable in the two groups is the same; it is possible
for the outcome data in the two groups to have similar measures of central tendency
or location, such as mean and medians, but different distributions. Consider for
example two groups of size 50; group A has 48 observations with value 0 and 2 with
value 1 whilst group B has 26 observations with value 0 and 24 with a value of 2.
Both groups have a median value of 0 but the p-value from the Mann-Whitney U test
is < 0.001, indicating that the distribution of data in two groups is different.

The Mann-Whitney U test requires all the observations (for both groups combined) to
be ranked as if they were from a single sample. From this the test statistic U is
calculated; it is the number of all possible pairs of observations comprising one
observation from each sample for which the rank of value in the first group precedes
the rank of the value in the second group. This test statistic is then used to obtain a P
value.

The principle is best illustrated with a simple example. Consider the following two
samples of size six X=(0,6,5,1,1,6) and Y=(9,4,7,8,3,5). These are then ranked in
order as if they were from the same sample (the values for sample X are given in
bold):

Values 0 1 1 3 4 5 5 6 6 7 8 9
Ranks 1 2.5 2.5 4 5 6.5 6.5 8.5 8.5 10 11 12

Page 49
Having ranked the values altogether, these ranks are then added up separately for
each sample to get two separate totals (U statistics), U x=29.5 and Uy=48.5. A useful
check is that the sum of the ranks should add to n(n+1)/2. In this case
n(n+1)=12(12+1)/5=78. The smaller of the two U statistics is used obtain a P-value;
thus the value of U used for this example is 29.5. As with the t statistic above this
value is compared to tabulated critical values under the null hypothesis (table 1) to
obtain a P-value. Rank totals greater than the tabulated critical values are not
significant. In this case n1 and n2 are both 6 and the tabulated critical value is 26. As
the value of 29.5 is greater than this, the results do not reach statistical significance
at the 5% level, and there that there is insufficient evidence to reject the null that the
two groups differ in terms of the distribution of their data.

Table 1: Mann-Whitney test on unpaired samples: 5% levels of P (taken from

Swinscow and Campbell(Swinscow & Campbell 2002))

n1→ 2 3 4 5 6 7 8 9 10 11 12 13 14 15
n2
↓
4 10
5 6 11 17
6 7 12 18 26
7 7 13 20 27 36
8 3 8 14 21 29 38 49
9 3 8 15 22 31 40 51 63
10 3 9 15 23 32 42 53 65 78
11 4 9 16 24 34 44 55 68 81 96
12 4 10 17 26 35 46 58 71 85 99 115
13 4 10 18 27 37 48 60 73 88 103 119 137
14 4 11 19 28 38 50 63 76 91 106 123 141 160
15 4 11 20 29 40 52 65 79 94 110 127 145 164 185
16 4 12 21 31 42 54 67 82 97 114 131 150 169
17 5 12 21 32 43 56 70 84 100 117 135 154
18 5 13 22 33 45 58 72 87 103 121 139
19 5 13 23 34 46 60 74 90 107 124
20 5 14 24 35 48 62 77 93 110
21 6 14 25 37 50 64 79 95
22 6 15 26 38 51 66 82
23 6 15 27 39 53 68
24 6 16 28 40 55
25 6 16 28 42
26 7 17 29
27 7 17
28 7

The previous tutorial illustrated the use of the independent samples t-test with some
data taken from a community leg ulcer trial(Morrell, Walters, Dixon, Collins, Brereton,
Peters, & Brooker 1998). For the leg ulcer data, there were 120 patients in the clinic
group and their mean number of ulcer free weeks for was 20.1. There were 113
patients in the control group and they had a mean number of ulcer free weeks of
14.2. It was demonstrated that there was a statistically significant difference in the

Page 50
number of ulcer free weeks between the two groups (P=0.014). However, if the
number of ulcer free weeks in each group is plotted it can be seen that the data are
highly skewed and are not Normally distributed (Figure 1a and 1b).

Figure 1a: Ulcer-free time for clinic Figure 1b: Ulcer-free time for home
group group
60 60

50 50

40 40
Frequency

Frequency
30 30

20 20

10 10

0 0
0.00 10.00 20.00 30.00 40.00 50.00
0.00 10.00 20.00 30.00 40.00 50.00
Leg ulcer free time (weeks) Leg ulcer free time (weeks)

If, instead of the independent samples t-test a Mann-Whitney U test were carried out
on these data the P-value obtained would be 0.017, a value that is remarkably similar
to that obtained from the t-test. In fact, the t-test and the Mann-Whitney U test will
tend to give similar P-values when the samples are large and approximately equal in
size). As this is less than the nominal level usually set for statistical significance of
0.05 we can reject the null hypothesis (that the distribution of the data in the two
groups are the same). We conclude that the result is statistically significant and there
is evidence that the distribution of ulcer free weeks is different between the two
groups. However, we are unable to state what the difference might be, only that there
is a difference, if we only consider the P-value.

Two groups of paired observations

When there is more than one group of observations it is vital to distinguish the case
where the data are paired from that where the groups are independent. Paired data
may arise when the same individuals are studied more than once, usually in different
circumstances, or when individuals are paired as in a case-control study. For
example, as part of the leg ulcer trial, data were collected on health related quality of
life (HRQoL) at baseline, 3 months and 12 months follow-up. The previous tutorial
described a method for analysing paired continuous data, the paired t-test.

If the assumptions underlying the use of the paired t-test are not met a non-
parametric alternative, the Wilcoxon signed rank sum test, can be used, This test is
based upon the ranks of the paired differences and the null hypothesis is that there is
no tendency for the outcome in one group (or under one condition) to be higher or
lower than in the other group (or condition). It assumes that (a) the paired differences
were independent of each other and (b) the differences come from a symmetrical
distribution (this can be checked by eye). As with the Mann-Whitney U test outlined
above the Wilcoxon signed rank sum test is most easily illustrated using an example.
Swinscow and Campbell (Swinscow & Campbell 2002) give details of a study of

Page 51
foetal movements before and after chorionic villus sampling. The data are shown in
table 2:

Table 2 Wilcoxon test on percentage of time foetus spent moving before and
after chononic villus sampling for ten pregnant women (Boogert et al. 1987)

Patient no Before After Difference Rank Signed

Sampling Sampling (before- rank
(2) (3) after) (4) (5) (6)
1 25 18 7 9 9
2 24 27 -3 5.5 -5.5
3 28 25 3 5.5 5.5
4 15 20 -5 8 -8
5 20 17 3 5.5 5.5
6 23 24 -1 1.5 -1.5
7 21 24 -3 5.5 -5.5
8 20 22 -2 3 -3
9 20 19 1 1.5 1.5
10 27 19 8 10 10

The differences between before and after sampling are calculated (colum 4) and
these are then ranked by size irrespective of sign (column 5; zero values omitted).
When two or more differences are identical each is allotted the point half way
between the ranks they would fill if distinct, irrespective of the plus or minus sign. For
instance, the differences of –1 (patient 6) and +1 (patient 9) fill ranks 1 and 2. As (1 +
2)/2 = 1.5, they are allotted rank 1.5. In column (6) the ranks are repeated for column
(5), but to each is attached the sign of the difference from column (4). A useful check
is that the sum of the ranks must add to n(n + 1)/2. In this case 10(10 + 1)/2 = 55.

The numbers representing the positive ranks and the negative ranks in column (6)
are added up separately and only the smaller of the two totals (irrespective of its
sign) is used to obtain a P-value from tabulated critical values under the null
hypothesis (Table 3). As with the Mann-Whitney U test rank totals greater then the
tabulated critical value are non-significant at the 5% level. In this case the smaller of
the two ranks is 23.5 and as this is larger than the number given for ten pairs in table
3 the result is not statistically significant. There is insufficient evidence to reject the
null that the median difference in foetal movements before and after sampling is zero.
We can conclude that we have little evidence that chorionic villus sampling alters the
movement of the foetus.

Page 52
Table 3: Wilcoxon test on paired samples: 5% and 1% levels of P (taken from
Swinscow and Campbell(Swinscow & Campbell 2002))

Number of 5% level 1%
pairs level
7 2 0
8 2 0
9 6 2
10 8 3
11 11 5
12 14 7
13 17 10
14 21 13
15 25 16
16 30 19

Note, perhaps contrary to intuition, that the Wilcoxon test, although a test based on
the ranks of the data values, may give a different value if the data are transformed,
say by taking logarithms. Thus it may be worth plotting the distribution of the
differences for a number of transformations to see if they make the distribution
appear more symmetrical.

Summary

Outlined above are some non-parametric methods for comparing two groups of
continuous data when the assumptions underlying the t-test (paired and unpaired)
are not met. However, as stated in the previous tutorial statistical significance does
not necessarily mean the result obtained is clinically significant or of any practical
importance. A P value will only indicate how likely the results obtained are when the
null hypothesis is true. Much more information, such as whether the result is likely to
be of clinical importance can be gained by calculating a confidence interval, as this a
range of plausible values for the estimated quantity. Details of how to do this can be
found in Statistics with Confidence(Altman et al. 2000)

Page 53
Page 54
The analysis of categorical data
Jenny V Freeman, Steven A Julious

Introduction

In the previous two tutorials we have outlined several methods for analysing potential
differences between two independent groups for a continuous outcome variable
(Freeman and Campbell 2006;Freeman & Julious 2006a). In this tutorial we will
discuss several methods for analysing differences between two independent groups
when the outcome of interest is binary; a comparison of the two proportions using the
Normal approximation to the Binomial and the Chi-squared test. If you recall from the
earliest tutorial(Freeman & Julious 2005c), binary data are data which can take only
two possible values such as healed / not healed or pregnant/ not pregnant.

The type of statistical test depends upon the answers to five key questions, as
outlined in tutorial six (Freeman & Julious 2006a). Briefly, these are:
The aims and objectives of the study
The hypothesis to be tested
The type of outcome data
The distribution of the outcome data
The appropriate summary measure for the outcome data.

Example data

The data in this tutorial have come from a randomised controlled trial of community
leg ulcer clinics(Morrell, Walters, Dixon, Collins, Brereton, Peters, & Brooker 1998).
Patients were allocated to one of two groups; either usual care by the district nursing
team (control), or care in a specialist leg ulcer clinic (clinic). The aim of the study was
to compare the treatments. One of the outcomes from the trial was whether the leg
ulcer was healed at 12 weeks and thus the null hypothesis is that there is no
difference between the two groups with respect to leg ulcer healing. The alternative
hypothesis is that there is a difference between the two groups.

In this case the outcome is binary (healed/ not healed) and the data are considered
to have a binomial distribution. The comparison is between two independent groups
(control and clinic groups). Of the 120 patients in the clinic group, 22 (18%) had a
healed leg ulcer at 12 weeks and of the 113 patients in the control group, 17 (15%)
had a healed leg ulcer at 12 weeks.

These data consist of the frequencies in each group with or without healed ulcers
and the simplest way to present them is in the form of a table such as table 1 below.
This is known as a 2 by 2 contingency table, because there are two rows and two
columns and it is said to have 4 cells (2 x 2). Generally a contingency with r rows and
c columns is known as an r by c contingency table and has r x c cells. For example, if
there were three treatment groups (rather than the two here) Table 1 would have 3
columns for the groups rather than two and thus the table would be a 2 by 3
contingency table (with 6 cells).

Page 55
Table 1: 2 x 2 contingency table of Treatment (clinic/home) by Outcome (ulcer
healed / not healed) for the Leg ulcer study data2
Treatment
Clinic Home Total
Outcome:
Healed 22 (18%) 17 (15%) 39
Not healed 98 (82%) 96 (85%) 194
Total 120 (100%) 113 (100%) 233

Comparison of two proportions

There are several approaches to analyse these data. One of the simplest is a
comparison of the proportion healed between the two groups (technically this is
known as the Normal approximation to the binomial distribution). In this case the
hypothesis test assumes that there is a common proportion (of healed ulcers), π,
which is estimated by p where p is the proportion of the total for both groups with a
healed leg ulcer. Consider the general form of the table above:

Table 2: General form of 2 x 2 table

Treatment group
1 2 Total
Outcome:
1 a (=n1p1) c (=n2p2) a+c (=np)
2 b (=n1(1-p1)) d (=n2(1-p2)) b+d (=n(1-p))
Total a+b (=n1) c+d (=n2) a+b+c+d (= n)

The common proportion p = a+c / (a+b+c+d)

If we let

p1 = a / (a+b) and p2 = c / (c+d)

the common proportion, p, can also be written as:

n1 p1 n2 p2
n1 n2

and the standard error for the difference in proportions is estimated by:

1 1
se( p1 p2 ) p(1 p)
n1 n2

From this we can compute the test statistic

p1 p2
z ,
se( p1 p2 )

2
When organising data such as this is it good practice to arrange the table with the grouping
variable forming the columns and the outcome variable forming the rows.

Page 56
which, under the null hypothesis (of no difference) is assumed to be Normally
distributed, with a mean of 0 and a standard deviation of 1 (i.e. a standard Normal
distribution). The reason we can make this Normal approximation is due to the
Central Limit theorem discussed in a previous note(Freeman & Julious 2005b). We
can then compare this value to what would be expected under the null hypothesis of
no difference, in order to get a P-value.

For the example above

(120 * 0.183) (113* 0.150) 38.91

p 0.167
120 113 233

1 1
se( p1 p2 ) 0.167 (1 0.167 ) 0.049
120 113

0.033
and thus z 0.673
0.049

Comparing this z statistic value to the value expected under the null hypothesis gives
a P-value of 0.502 (i.e. the probability of observing z=0.67 or a value more extreme if
the null hypothesis is true is 0.502). As this is greater than 0.05, we are unable to
reject to null and we would conclude that there is no reliable evidence of a difference
in leg ulcer healing rates at 12 weeks between the clinic and control groups. In
addition to carrying out a significance test we could also calculate a confidence
interval for the difference in proportions.

The 95% confidence interval for the difference in proportions is given by

( p1 p2 ) 1.96 se( p1 p2 ) .

Using the data from the example above, the 95% confidence interval for the true
difference in leg ulcer healing rates between the two groups is thus 0.063 to 0.129.
Therefore we are 95% confident that the true population difference in the proportion
of leg ulcers healed, at 12 weeks, between the clinic and control groups lies between
-0.063 and 0.128 and the best estimate is 0.033. These may also be expressed as
percentages, such that the 95% confidence interval is given by -0.63% and 12.8%.

This approach is only valid when the sample is large enough for the Normal
approximation to the binomial to apply; as a rule of thumb both np and n(1-p) should
exceed 5 where n=total number of individuals in both groups, p=proportion of
individuals with the outcome of interest (irrespective of group) and (1-p) is the
proportion of individuals without the outcome of interest (irrespective of group) (see
table 3). In addition, thinking of it as a difference in two proportions only makes sense
for 2x2 tables i.e. where there are only two groups and two outcomes.

Page 57
Table 3: Assumptions for the Normal approximation to the Binomial to be valid:

Both np and n(1-p) must be > 5

Where n = total number of individuals in both samples

p = proportion of individuals with condition (irrespective of group)
(1-p) = proportion of individuals without condition (irrespective of group)

Chi-squared test

An alternative approach to analysing the data contained in table 1 and by far the
most common, is to apply the Chi-squared test (χ2 test). This is a more general test
and may be used when there are two independent unordered categorical variables
that form an r x c contingency table (n.b the current example is a 2 x 2 table, where
both r and c = 2). It is valid when at least 80% of expected cell frequencies are
greater than 5 and all expected cell frequencies ≥1.

Table 4: Assumptions for the Chi-squared test to be valid:

The two variables are independent

At leat 80% of expected cell frequencies are > 5
All expected cell frequencies are ≥1

The null hypothesis for the chi-squared test is that there is no relationship between
the row variable and the column variable, i.e. that being in a particular column does
not influence whether you will be in a particular row, and visa versa. For the example
above the null hypothesis is that both treatments have the same effect i.e. that being
in a particular treatment group (column) is unrelated to whether the leg ulcer healed
(row).

If the null hypothesis were true we would expect both treatments to have the same
effect and the same the proportion of leg ulcers healed in each group. We estimate
this proportion based on the overall proportion healed in both groups. Thus the best
estimate of the common “ulcer healing” rate at 12 weeks is given as the common
proportion 39 /233 = 16.7% and we use this to calculate the expected number healed
for each group. Under the null hypothesis, if there was no relationship between the
study group and outcome, we would expect 16.7% of leg ulcers to be healed in each
group. For example there are 120 patients in the clinic group and we would expect
20.1 of them to have a healed ulcer by 12 weeks (16.7% of 120). We can calculate
the expected number (frequency) in each of the four cells as the (column total*row
total)/ overall total. The steps to calculate the Chi-squared statistic are outlined in the
table below

Page 58
Steps to calculate the Chi-squared test statistic

1. Calculate the expected value for each of the four cells in the table
2. Calculate the difference between the observed value and the expected value
for each cell
3. Square each difference and divide the resultant quantity by the expected value
4. Add all of these values to get a single number, the χ 2 statistic
5. Calculate the degrees of freedom (df): (number of rows – 1) x (number of
columns – 1)
6. Compare this number with tabulated values of the χ2 distribution with the same
degrees of freedom as calculated above

For the leg ulcer example the χ2 statistic is 0.445 and there is 1 degree of freedom.
Comparing this to tabulated values of the χ2 distribution with 1df gives a P-value of
0.502. As this is greater than the nominal significance level of 0.05 the result is said
to be not statistically significant. From this there is insufficient evidence to reject the
null and we can conclude that there is no reliable evidence of a difference in leg ulcer
healing rates between the two groups.

Chi-squared with continuity correction

In 2 x 2 tables, even when the expected cell counts are bigger than 5, the
mathematical approximations for the test statistics are sub-optimal and the null
hypothesis is rejected too often on average. In order to overcome this, a modification
has been suggested to the formula for the chi-squared test, known as Yates‟
continuity correction and it is recommended for all 2 x 2 tables (Altman 1991). This
continuity correction involves adding on 0.5 to each individual cell‟s contribution to
the overall chi-squared such that the formula for the overall chi-squared may be
written
2
2 O E 0.5
cc
E
and this can again be compared with tables for the chi-squared distribution on 1 df.

For the leg ulcer data:

O E |O-E| - 0.5 (|O-E| - 0.5)2 |O-E| - 0.52 / E
Healed / clinic 22 20.1 1.4 1.96 0.98
Not healed / 98 99.9 1.4 1.96 0.020
clinic 17 18.9 1.4 1.96 0.104
Healed / 96 94.1 1.4 1.96 0.021
control
Not healed /
control
Total 233 233 0.243

Thus the cc2 = 0.243 and under the null hypothesis of no association between the
rows and columns the probability of observing this value of the test statistic or more
extreme is about 0.62. Note that this P-value is greater than that obtained without the
continuity correction, as this test corrects for the fact that on average the null
hypothesis will be rejected to often. This test is said to be more conservative than the

Page 59
original one. The advantage of the continuity correction is that it is easy to implement.
It was more commonly used in the past than today where computer intensive
methods such as Fisher‟s Exact test can be readily applied.

Fisher’s exact test

In a 2 x 2 table when the expected cell counts are smaller than 5 or any are less than
1 even Yates‟ correction does not work. In this case Fisher‟s Exact test, proposed by
RA Fisher, can be applied. The test is based upon calculating the probability of
rejecting the null hypothesis directly, using all possible tables that could have been
observed. This will not be dealt with in more detail here as it will from the basis of a
later note. It is mentioned here merely as an alternative test when the assumption
underlying the chi-squared test are found not to be valid.

Summary

This tutorial has dealt with some simple methods for analysing binary data. It has
outlined how such data can be tabulated using contingency tables and how these
can be analysed. Provided the number of observations is large enough, the
proportions in the two groups can be compared directly. An alternative is to use a
more general test, called the chi-squared test, and this can be extended to more than
two groups, or more than two possible outcomes. However, when there are only two
groups and two outcomes it is recommended that Yates‟ continuity correction is used
as the standard, or where the assumptions are not valid, Fisher‟s Exact test.

Page 60
The analysis of categorical data: Fisher’s Exact test
Jenny V Freeman, Michael J Campbell

Introduction

In the previous tutorial we have outlined some simple methods for analysing binary
data, including the comparison of two proportions using the Normal approximation to
the binomial and the Chi-squared test((Freeman and Julious 2007)). However, these
methods are only approximations, although the approximations are good when the
sample size is large. When the sample size is small we can evaluate all possible
combinations of the data and compute what are known as exact P-values.

Fisher’s Exact test

When one of the expected values (note: not the observed values) in a 2x2 table is
less than 5, and especially when it is less than 1, then Yates‟ correction can be
improved upon. In this case Fisher‟s Exact test, proposed in the mid-1930s almost
simultaneously by Fisher, Irwin and Yates(Armitage et al. 2002), can be applied. The
null hypothesis for the test is that there is no association between the rows and
columns of the 2x2 table, such that the probability of a subject being in a particular
row is not influenced by being in a particular column. If the columns represented the
study group and the rows represented the outcome, then the null hypothesis could be
interpreted as the probability of having a particular outcome is not influenced by the
study group, and the test evaluates whether the two study groups differ in the
proportions with each outcome.

An important assumption for all of the methods outlined, including Fisher‟s exact test,
is that the binary data are independent. If the proportions are correlated, then more
advanced techniques should be applied. For example in the leg ulcer example of the
previous tutorial(Freeman & Julious 2007), if there were more than one leg ulcer per
patient, we could not treat the outcomes as independent.

The test is based upon calculating directly the probability of obtaining the results that
we have obtained (or results more extreme) if the null hypothesis is actually true,
using all possible 2x2 tables that could have been observed, for the same row and
column totals as the observed data. These row and column totals are also known as
marginal totals. What we are trying to establish is how extreme our particular table
(combination of cell frequencies) is in relation all the possible ones that could have
occurred given the marginal totals.

This is best explained by a simple worked example. The data below come from an
RCT comparing intra-muscular magnesium injections with placebo for the treatment
of chronic fatigue syndrome(Cox et al. 1991). Of the 15 patients who had the intra-
muscular magnesium injections 12 felt better (80%), whereas, of the 17 on placebo,
only 3 felt better (18%).

Page 61
Table 1: Results of the study to examine whether intramuscular magnesium is
better than placebo for the treatment of chronic fatigue syndrome 3
Magnesium Placebo Total
Felt better 12 3 15
Did not feel 3 14 17
better
Total 15 17 32

There are 16 different ways of rearranging the cell frequencies for the above table,
whilst keeping the marginal totals the same, as illustrated below in figure 1. The
result that corresponds to our observed cell frequencies is (xiii):

Figure 1: Illustration of all the different ways of rearranging cell frequencies in

table 1, but with the marginal totals remaining the same

(i) 0 15 (ii) 1 14 (iii) 2 13 (iv) 3 12

15 2 14 3 13 4 12 5

(v) 4 11 (vi) 5 10 (vii) 6 9 (viii) 7 8

11 6 10 7 9 8 8 9

(ix) 8 7 (x) 9 6 (xi) 10 5 (xii) 11 4

7 10 6 11 5 12 4 13

(xiii) 12 3 (xiv) 13 2 (xv) 14 1 (xvi) 15 0

3 14 2 15 1 16 0 17

The general form of table 1 is given in table 2 and under the null hypothesis of no
association Fisher showed that the probability of obtaining the frequencies, a, b, c
and d in table 2 is

(a b)!(c d )!(a c)!(b d )!

(1)
(a b c d )!a!b!c!d!

where x! is the product of all the integers between 1 and x, e.g. 5!=1x2x3x4x5=120
(note that for the purpose of this calculation, we define 0! as 1). Thus for each of the
results (i) to (xvi) the exact probability of obtaining that result can be calculated (table
3). For example, the probability of obtaining table (i) in figure 1 is
15!17!15!17! =0.0000002
32!0!15!15!2!

3
When organising data such as this is it good practice to arrange the table with the grouping
variable forming the columns and the outcome variable forming the rows.

Page 62
Table2: General form of table 1
Column 1 Column 2 Total
Row 1 a b a+b
Row 2 c d c+d
Total a+c b+d a+b+c+d

Table 3: Probabilities associated with each of the frequency tables above,

calculated using formula 1
Table a b c d P-value
i 0 15 15 2 0.0000002
ii 1 14 14 3 0.0000180
iii 2 13 13 4 0.0004417
iv 3 12 12 5 0.0049769
v 4 11 11 6 0.0298613
vi 5 10 10 7 0.1032349
vii 6 9 9 8 0.2150728
viii 7 8 8 9 0.2765221
ix 8 7 7 10 0.2212177
x 9 6 6 11 0.1094916
xi 10 5 5 12 0.0328475
xii 11 4 4 13 0.0057426
xiii 12 3 3 14 0.0005469
xiv 13 2 2 15 0.0000252
xv 14 1 1 16 0.0000005
xvi 15 0 0 17 0.0000000

From table 3 we can see that the probability of obtaining the observed frequencies
for our data is that which corresponds with (xiii), which gives p=0.0005469 and the
probability of obtaining our results or results more extreme (a difference that is at
least as large) is the sum of the probabilities for (xiii) to (xvi) = 0.000573. This gives
the one-sided P-value or obtaining our results or results more extreme, and in order
to obtain the two-sided p-value there are several approaches. The first is to simply
double this value, which gives p=0.0001146. A second approach is to add together
all the probabilities that are the same size or smaller than the one for our particular
result, in this case, all probabilities that are less than or equal to 0.0005469, which
are tables (i), (ii), (iii), (xiii), (xiv), (xv) and (xvi). This gives a two-sided value of
p=0.001033. Generally the difference is not great, though the first approach will
always give a value greater than the second. A third approach, which is
recommended by Swinscow and Campbell (Swinscow & Campbell 2002) is a
compromise and is known as the mid-p method. All the values more extreme than
the observed p-value are added up and these are added to one half of the observed
value. This gives p=0.000759.

The criticism of the first two methods is that they are too conservative, i.e. is the null
hypothesis was true, over repeated studies they would reject the null hypothesis less
often than 5%. They are conditional on both sets of marginal totals being fixed, i.e.
exactly 15 people being treated with magnesium and 15 feeling better. However, if
the study were repeated, even with 15 and 17 in the magnesium and placebo groups

Page 63
respectively, we would not necessarily expect exactly 15 to feel better. The mid-p
value method is less conservative, and gives approximately the correct rate of type I
errors (false positives).

In either case, for our example, the P-value is less than 0.05, the nominal level for
statistical significance and we can conclude that there is evidence of a statistically
significant difference in the proportions feeling better between the two treatment
groups. However, in common with other non-parametric tests, Fisher‟s exact test is
simply a hypothesis test. It will merely tell you whether a difference is likely, given the
null hypothesis (of no difference). It gives you no information about the likely size of
the difference, and so whilst we can conclude that there is a significant difference
between the two treatments with respect to feeling better or not, we can draw no
conclusions about the possible size of the difference.

Example data from last week

Table 1 shows the data from the previous tutorial. It is from a randomised controlled
trial of community leg ulcer clinics(Morrell, Walters, Dixon, Collins, Brereton, Peters,
& Brooker 1998), comparing the cost-effectiveness of community leg ulcer clinics
with standard nursing care. The columns represent the two treatment groups,
specialist leg ulcer clinic (clinic) and standard care (home), and the rows represent
the outcome variable, in this case whether the leg ulcer has healed or not.

Table 1: 2 x 2 contingency table of Treatment (clinic/home) by Outcome (ulcer

healed / not healed) for the Leg ulcer study
Treatment
Clinic Home Total
Outcome:
Healed 22 (18%) 17 (15%) 39
Not healed 98 (82%) 96 (85%) 194
Total 120 (100%) 113 (100%) 233

For this example the two-sided p-value from Fisher‟s exact test is 0.599 two-sided
and in this case we would not reject the null and would conclude that there is
insufficient evidence to

Summary

This tutorial has described in detail Fisher‟s exact test, for analysing simple 2x2
contingency tables when the assumptions for the chi-squared test are not met. It is
tedious to do by hand, but nowadays is easily computed by most statistical
packages.

Page 64
Use of Statistical Tables
Lucy Radford, Jenny V Freeman, Stephen J Walters

Introduction

Previous tutorials have looked at hypothesis testing(Freeman & Julious 2006b)and

basic statistical tests(Freeman & Campbell 2006;Freeman & Julious 2006a;Freeman
& Julious 2007). As part of the process of statistical hypothesis testing, a test statistic
is calculated and compared to a hypothesized critical value and this is used to obtain
a P-value. This P-value is then used to decide whether the study results are
statistically significant or not. This tutorial will explain how statistical tables are used
to link test statistics to P-values. It introduces tables for three important statistical
distributions: the standard Normal, t and chi-squared distributions and explains how
to use them with the help of some simple examples.

Standard Normal Distribution

The Normal distribution is widely used in Statistics and has been discussed in detail
previously(Freeman & Julious 2005b). As the mean of a Normally distributed variable
can take any value (-∞ to ∞) and the standard deviation any positive value (0 to ∞),
there are an infinite number of possible Normal distributions. It is therefore not
feasible to print tables for each Normal distribution; however it is possible to convert
any Normal distribution to the standard Normal distribution, for which tables are
available. The standard Normal distribution has a mean of 0 and standard deviation
of 1.

Any value X from a Normal distribution with mean μ and standard deviation σ can be
transformed to the standard Normal distribution using the following formula:

X
(1) z

This transformed X-value, often called z or z-score, is also known as the standard
Normal deviate, or Normal score. If an average, rather than a single value, is used
the standard deviation should be divided by the square root of the sample size, n, as
shown in equation (2).

X
(2) z
n

For example, the exam results for the first year of a medical degree are known to be
approximately Normally distributed with mean 72 and standard deviation 8. To find
the probability that a student will score 89 or more we first need to convert this value
to a standard Normal deviate. In this instance, as we have a single value we use
equation (1):

89 72
z 2.13
8

Page 65
If we wished to find the probability that an average of 10 scores is 75 or more we
would use equation (2) to convert to the standard Normal distribution:

75 72
z 1.19
8 10

We then use the standard Normal table to find the probabilities of observing these z
values, or values more extreme given that the population mean and standard
deviation are 72 and 8 respectively.

Standard Normal tables can be either one-tailed or two-tailed. In the majority of

hypothesis tests the direction of the difference is not specified, leading to a two-sided
(or two-tailed) test(Freeman & Julious 2006b). The standard Normal table shown in
Table 1 is two-sided4. In this two-sided table the value tabulated is the probability, α,
that a random variable, Normally distributed with mean zero and standard deviation
one, will be either greater than z or less than –z (as shown in the diagram at the top
of the table). The total area under the curve represents the total probability space for
the standard Normal distribution and sums to 1, and the shaded areas at either end
are equal to α/2. A one-tailed probability can be calculated by halving the tabulated
probabilities in Table 1. As the Normal distribution is symmetrical it is not necessary
for tables to include the probabilities for both positive and negative z values.

4
A simple trick for seeing whether a particular table is one-tailed or two-tailed is to look at the value that
corresponds to a cut-off of 1.96. If the tabulated P-value is 0.05 then the table is for two-tailed p-values.

Page 66
Table 1. Extract from two-tailed standard Normal table. Values tabulated are P-
values corresponding to particular cut-offs and are for z values calculated to
two decimal places.

z 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.00 1.0000 0.9920 0.9840 0.9761 0.9681 0.9601 0.9522 0.9442 0.9362 0.9283
0.10 0.9203 0.9124 0.9045 0.8966 0.8887 0.8808 0.8729 0.8650 0.8572 0.8493
0.20 0.8415 0.8337 0.8259 0.8181 0.8103 0.8206 0.7949 0.7872 0.7795 0.7718
0.30 0.7642 0.7566 0.7490 0.7414 0.7339 0.7263 0.7188 0.7114 0.7039 0.6965
0.40 0.6892 0.6818 0.6745 0.6672 0.6599 0.6527 0.6455 0.6384 0.6312 0.6241
0.50 0.6171 0.6101 0.6031 0.5961 0.5892 0.5823 0.5755 0.5687 0.5619 0.5552
0.60 0.5485 0.5419 0.5353 0.5287 0.5222 0.5157 0.5093 0.5029 0.4965 0.4902
0.70 0.4839 0.4777 0.4715 0.4654 0.4593 0.4533 0.4473 0.4413 0.4354 0.4295
0.80 0.4237 0.4179 0.4122 0.4065 0.4009 0.3953 0.3898 0.3843 0.3789 0.3735
0.90 0.3681 0.3628 0.3576 0.3524 0.3472 0.3421 0.3371 0.3320 0.3271 0.3222
1.00 0.3173 0.3125 0.3077 0.3030 0.2983 0.2837 0.2891 0.2846 0.2801 0.2757
1.10 0.2713 0.2670 0.2627 0.2585 0.2543 0.2501 0.2460 0.2420 0.2380 0.2340
1.20 0.2301 0.2263 0.2225 0.2187 0.2150 0.2113 0.2077 0.2041 0.2005 0.1971
1.30 0.1936 0.1902 0.1868 0.1835 0.1802 0.1770 0.1738 0.1707 0.1676 0.1645
1.40 0.1615 0.1585 0.1556 0.1527 0.1499 0.1471 0.1443 0.1416 0.1389 0.1362
1.50 0.1336 0.1310 0.1285 0.1260 0.1236 0.1211 0.1188 0.1164 0.1141 0.1118
1.60 0.1096 0.1074 0.1052 0.1031 0.1010 0.0989 0.0969 0.0949 0.0930 0.0910
1.70 0.0891 0.0873 0.0854 0.0836 0.0819 0.0801 0.0784 0.0767 0.0751 0.0735
1.80 0.0719 0.0703 0.0688 0.0672 0.0658 0.0643 0.0629 0.0615 0.0601 0.0588
1.90 0.0574 0.0561 0.0549 0.0536 0.0524 0.0512 0.0500 0.0488 0.0477 0.0466
2.00 0.0455 0.0444 0.0434 0.0424 0.0414 0.0404 0.0394 0.0385 0.0375 0.0366
2.10 0.0357 0.0349 0.0340 0.0332 0.0324 0.0316 0.0308 0.0300 0.0293 0.0285
2.20 0.0278 0.0271 0.0264 0.0257 0.0251 0.0244 0.0238 0.0232 0.0226 0.0220
2.30 0.0214 0.0209 0.0203 0.0198 0.0193 0.0188 0.0183 0.0178 0.0173 0.0168
2.40 0.0164 0.0160 0.0155 0.0151 0.0147 0.0143 0.0139 0.0135 0.0131 0.0128
2.50 0.0124 0.0121 0.0117 0.0114 0.0111 0.0108 0.0105 0.0102 0.0099 0.0096
2.60 0.0093 0.0091 0.0088 0.0085 0.0083 0.0080 0.0078 0.0076 0.0074 0.0071
2.70 0.0069 0.0067 0.0065 0.0063 0.0061 0.0060 0.0058 0.0056 0.0054 0.0053
2.80 0.0051 0.0050 0.0048 0.0047 0.0045 0.0044 0.0042 0.0041 0.0040 0.0039
2.90 0.0037 0.0036 0.0035 0.0034 0.0033 0.0032 0.0031 0.0030 0.0029 0.0028
3.00 0.0027 0.0026 0.0025 0.0024 0.0024 0.0023 0.0022 0.0021 0.0021 0.0020

Page 67
From our first example above we want to know what the probability is that a student
chosen at random will have a test score of 89, given a population mean of 72 and
standard deviation of 8. The z-score calculated above is 2.13. In order to obtain the
P-value that corresponds to this z-score we first look at the row in the table that
corresponds to a z-score of 2.1. We then need to look down the column that is
headed 0.03. The corresponding P-value is 0.0198. However, this is a two-sided
probability and corresponds to probability that a z-score is either -2.13 or 2.13 (see
figure 1). To get the probability that a student chosen at random will have a test score
of at least 89 we need to halve the tabulated P-value. This gives a P-value of 0.0099.

Figure 1: Normal curve showing the Z values and corresponding P-values for
the data in example 1.

In a previous tutorial we used the Normal approximation to the binomial to examine

whether there were significant differences in the proportion of patients with healed
leg ulcers at 12 weeks, between standard treatment and treatment in a specialised
leg ulcer clinic(Freeman & Julious 2007). The null hypothesis was that there was no
difference in healing rates between the two groups. From this test we obtained a z
score of 0.673. Looking this up in Table 1 we can see that it corresponds to a two-
sided P-value of 0.503. Thus we cannot reject the null, and we conclude that there is
no reliable evidence of a difference in ulcer healing rates at 12 weeks between the
two groups.

Student’s t-Distribution

The t-test is used for continuous data to compare differences in means between two
groups (either paired or unpaired)(Freeman & Julious 2006a). It is based on
Student‟s t-distribution (sometimes referred to as just the t-distribution). This
distribution is particularly important when we wish to estimate the mean (or mean

Page 68
difference between groups) of a Normally distributed population but have only a small
sample. This is because the t-test, based on the t-distribution, offers more precise
estimates for small sample sizes than the tests associated with the Normal
distribution. It is closely related to the Normal distribution and as the sample size
tends towards infinity the probabilities of the t-distribution approach those of the
standard Normal distribution.

The main difference between the t-distribution and the Normal distribution is that the t
depends only on one parameter, v, the degrees of freedom (d.f.), not on the mean or
standard deviation. The degrees of freedom are based on the sample size, n, and
are equal to n – 1. If the t statistic calculated in the test is greater than the critical
value for the chosen level of statistical significance (usually P = 0.05) the null
hypothesis for the particular test being carried out is rejected in favour of the
alternative. The critical value that is compared to the t statistic is taken from the table
of probabilities for the t-distribution, an extract of which is shown in Table 2.

Unlike the table for the Normal distribution described above the tabulated values
relate to particular levels of statistical significance, rather than the actual P-values.
Each of the columns represents the cut-off points for declaring statistical significance
for a given level of (two-sided) significance. For example, the column headed 0.05 in
Table 2 gives the values which a calculated t-statistic must be above in order for a
result to be statistically significant at the two-sided 5% level. Each row represents the
cut-offs for different degrees of freedom. Any test which results in a t statistic less
than the tabulated value will not be statistically significant at that level and the P-
value will be greater than the value indicated in the column heading. As the t-
distribution is symmetrical about the mean, it is not necessary for tables to include
the probabilities for both positive and negative t statistics.

Consider for example, a t-test from which a t value of 2.66 on 30 d.f was obtained.
Looking at the row corresponding to 30 d.f. in Table 2 this value falls between the
tabulated values for 0.02 (=2.457) and 0.01 (=2.75). Thus, the P-value that
corresponds with this particular t value will be less than 0.02, but greater than 0.01.
In fact the actual (two-tailed) P-value is 0.012.

Page 69
Table 2. Distribution of t (two-tailed) taken from Swinscow &
Campbell(Swinscow & Campbell 2002)
Probability
d.f. 0.5 0.1 0.05 0.02 0.01 0.00l

1 l.000 6.314 12.706 3l.821 63.657 636.6l9

2 0.816 2.920 4.303 6.965 9.925 31.598
3 0.765 2.353 3.182 4.541 5.841 12.941
4 0.741 2.132 2.776 3.747 4.604 8.610
5 0.727 2.015 2.571 3.365 4.032 6.859
6 0.718 1.943 2.447 3.l43 3.707 5.959
7 0.711 1.895 2.365 2.998 3.499 5.405
8 0.706 l.860 2.306 2.896 3.355 5.04l
9 0.703 l.833 2.262 2.82l 3.250 4.78l
10 0.700 l.812 2.228 2.764 3.169 4.587
11 0.697 1.796 2.201 2.718 3.l06 4.437
12 0.695 1.782 2.179 2.681 3.055 4.3l8
13 0.694 1.771 2.160 2.650 3.012 4.221
14 0.692 1.76l 2.145 2.624 2.977 4.l40
15 0.69l l.753 2.13l 2.602 2.947 4.073
16 0.690 1.746 2.120 2.583 2.92l 4.015
17 0.689 1.740 2.110 2.567 2.898 3.965
18 0.688 1.734 2.101 2.552 2.878 3.922
19 0.688 l.729 2.093 2.539 2.861 3.883
20 0.687 1.725 2.086 2.528 2.845 3.850
21 0.686 1.721 2.080 2.518 2.831 3.8l9
22 0.686 1.717 2.074 2.508 2.819 3.792
23 0.685 1.714 2.069 2.500 2.807 3.767
24 0.685 1.711 2.064 2.492 2.797 3.745
25 0.684 1.708 2.060 2.485 2.787 3.725
26 0.684 1.706 2.056 2.479 2.779 3.707
27 0.684 1.703 2.052 2.473 2.771 3.690
28 0.683 1.701 2.048 2.467 2.763 3.674
29 0.683 1.699 2.045 2.462 2.756 3.659
30 0.683 l.697 2.042 2.457 2.750 3.646
40 0.681 l.684 2.021 2.423 2.704 3.551
60 0.679 1.671 2.000 2.390 2.660 3.460
120 0.677 1.658 l.980 2.358 2.617 3.373
∞ 0.674 1.645 1.960 2.326 2.576 3.291

Chi-squared Distribution

The final statistical table being considered in this tutorial is that of the chi-squared
distribution. There are a wide range of statistical tests that lead to use of the chi-
squared distribution, the most common of which is the chi-squared test described in a
previous tutorial(Freeman & Julious 2007). Like the t-distribution the chi-squared
distribution has only one parameter, the degrees of freedom, k. A section of the chi-
squared distribution is shown in Table 3. Like the table for the t distribution described
above the tabulated values are the chi-squared values that relate to particular levels

Page 70
of statistical significance, rather than actual P-values. Each of the columns
represents the cut-off points for declaring statistical significance for a given level of
significance. For example, the column headed 0.05 in Table 3 gives the values above
which a calculated chi-squared statistic must be in order for a result to be statistically
significant at the two-sided 5% level, for degrees of freedom ranging from 1 to 30.
Any test which results in a chi-squared statistic less than the tabulated value will not
be statistically significant at that level and the P-value will be greater than the value
at the top of the column. Consider, for example, a chi-squared value of 4.2 on 1 d.f.
Looking at the row corresponding to 1 d.f. in Table 3 this value falls between the
tabulated values for 0.05 (=3.841) and 0.02 (=5.412). Thus, the P-value that
corresponds with this particular chi-squared statistic will be less than 0.05, but
greater than 0.02.

As a second example consider the results of a chi-squared test that was used to
assess whether leg ulcer healing rates differed between two different treatment
groups (group 1: standard care; treatment 2: specialised leg ulcer clinic)(Freeman &
Julious 2007). From this significance test a chi-squared value of 0.243 with 1 d.f. was
obtained. Looking at the 1 d.f. row in Table 3 it can be seen that all the values are
greater than this value, including the value that corresponds with a P-value of 0.5,
0.455. Thus we can conclude that the P-value corresponding to a chi-squared value
of 0.243 is greater than 0.5; in fact the exact value is 0.62.

Page 71
Table 3. Distribution of χ2 taken from Swinscow & Campbell(Swinscow &
Campbell 2002)

Probability*
d.f. 0.5 0.10 0.05 0.02 0.01 0.00l

1 0.455 2.706 3.841 5.412 6.635 10.827

2 1.386 4.605 5.991 7.824 9.210 13.815
3 2.366 6.251 7.815 9.837 11.345 16.268
4 3.357 7.779 9.488 11.668 13.277 18.465
5 4.351 9.236 11.070 13.388 15.086 20.517
6 5.348 10.645 12.592 15.033 16.812 22.457
7 6.346 12.017 14.067 16.622 18.475 24.322
8 7.344 l3.362 15.507 18.168 20.090 26.125
9 8.343 l4.684 16.919 19.679 21.666 27.877
10 9.342 l5.987 18.307 21.161 23.209 29.588
11 10.341 17.275 19.675 22.618 24.725 31.264
12 11.340 18.549 21.026 24.054 26.217 32.909
13 12.340 19.812 22.362 25.472 27.688 34.528
14 13.339 21.064 23.685 26.873 29.141 36.123
15 14.339 22.307 24.996 28.259 30.578 37.697
16 15.338 23.542 26.296 29.633 32.000 39.252
17 16.338 24.769 27.587 30.995 33.409 40.790
18 17.338 25.989 28.869 32.346 34.805 42.312
19 18.338 27.204 30.144 33.687 36.191 43.820
20 19.337 28.412 31.410 35.020 37.566 45.315
21 20.337 29.615 32.671 36.343 38.932 46.797
22 21.337 30.813 33.924 37.659 40.289 48.268
23 22.337 32.007 35.172 38.968 41.638 49.728
24 23.337 33.196 36.415 40.270 42.980 51.745
25 24.337 34.382 37.652 41.566 44.314 52.620
26 25.336 35.563 38.885 42.479 45.642 54.707
27 26.336 36.741 40.113 44.140 45.963 55.476
28 27.336 37.916 41.337 45.419 48.278 56.893
29 28.336 39.087 42.557 46.693 49.588 58.302
30 29.336 40.256 43.773 47.962 50.892 59.703
*: these are two-sided P-values

Summary

In this tutorial we have shown how to use statistical tables to obtain P-values for the
standard Normal, t and chi-squared distributions, and given examples to show how
the values from these tables are used to make decisions in a variety of basic
statistical tests.

Page 72
One-way Analysis of Variance
Jenny V Freeman, Michael J Campbell

In a previous tutorial we described the unpaired t-test for comparing two independent
groups when the data are Normally distributed(Freeman & Julious 2005a). In this
tutorial we will explain how this can be generalised to comparing more than two
groups using a method called the one-way analysis of variance (ANOVA). Examples
of such comparisons include:

1. Pain score between groups given different analgesics in a clinical trial

2. Birthweights between different methods of delivery for women in a particular
hospital
3. Heights of children with different ethnic backgrounds on entry to primary
school

Whilst it is possible to compare individual pairs of groups using the t-test this would
increase the probability of committing a type 1 error (false positive error) and is not
an efficient use of the data. The one-way ANOVA is a single global test of whether
the means differ in any of the groups. However, it is worth noting that if there are only
two groups then the one-way ANOVA is exactly equivalent to the t-test and will give
the same p-value.

There are several assumptions that need to be satisfied for the one-way ANOVA to
be valid as outlined in Box 1 and these should be checked before performing the test.
The assumption of equality of variances can be tested by examining the standard
deviations for each group, as a rule of thumb, no single SD should be greater than
twice any of the others. The assumption of normality can be examined by looking at
either dot plots of the data in groups or histograms, if the numbers are large. If the
assumptions are not met then the data can be transformed, for example by taking
logarithms, or by using the non-parametric equivalent of the ANOVA, the Krukall-
Wallis test. This test uses exactly the same methodology as the one-way ANOVA,
except that the data are ranked (ignoring grouping) and the test is performed on the
ranks.

Box 1: Assumptions underlying one-way ANOVA

1. The data are independent

2. The data are Normally distributed in each group
3. The variance is the same in each group

Description of technique

In brief, the one-way ANOVA is based upon the idea that you can partition the
variability in a set of data into different sources, for example into random variability
between individuals within groups (sometimes called the residual or unexplained
variability) and variability due to systematic difference between groups. Under the
null hypothesis that the means are the same, the within and between variances are
expected to be the same. However, if there are systematic differences between
groups then it would be expected that the between groups variance would be greater
than that within groups and a test can be constructed that is based upon the ratio of

Page 73
these two variances. This ratio is known as the F statistic and critical values for a
significance test can be obtained from tables of the F-distribution, but in order to do
this you need to know the degrees of freedom (df), of which there are two types.
There are those due to the variability between groups (df= number of groups -1) and
those due to the variability within groups (df= total number of observations – number
of groups).

In order the calculate the F-statistics, for each group you need to count the number of
observations (n), the mean for the variable of interest, y , the sum of the
observations(T), and the sum of the observations squared(S). Then sum each of
these quantities across the groups. Assume, for example, that you have k groups:

All groups
Group 1 2 … k combined
k
Number of observations n1 n2 … nk N ni
i 1
k
Sum of observations T1 T2 … Tk T Ti
i 1

Mean of observations y1 y 2 ... y3 y T N

k
Sum of squared observations S1 S2 … Sk S Si
i 1

Once the above quantities have been calculated you can then construct an analysis
of variance table to obtain the F statistic (Box 2):

Box 2: Analysis of Variance Table:

Sum of squares Degrees of Mean F statistic

(SS) freedom (df) Square (MS) (variance ratio)
k
2
Between groups Ti ni T2 N k-1 SS/df MSbetween/MSwithin
i 1

Within groups k N-k SS/df

S Ti 2 ni
i 1

Total N-1
S T2 N

This is then compared to tabulated critical values of an F statistic on k-1 and N-k
degrees of freedom to obtain a P-value.

Further details of the mathematics of the technique can be found in Chapter 7 of

Statistical Methods for Medical Research by Armitage et al(Armitage, Berry, &
Matthews 2002).

Page 74
Example

As part of a study looking at the post-operative pain experience of patients following

shoulder surgery, two different methods of providing surgical analgesia, subacromial
bursa block (SBB) and intrascalene block (ISB) were compared with a control
procedure(Nisar et al. 2007). There were 19 patients in each of the analgesia groups
and 9 in the control group and the outcome of interest was the amount of morphine
used by the patient in the first 24 hours after surgery. The data are displayed Figure
1. Examining this figure, it appears that there are systematic differences between the
study groups

Figure 1: Amount of patient controlled analgesia (mls morphine) used in first

24 hours after surgery by study group.

The null hypothesis is that there are no differences between the means of the three
groups, and this can be tested against the alternative that there are differences
between the means. Using the method outlined above the data can be summarised
as in table 1:

Page 75
Table 1: Construction of the components for the ANOVA table for the analgesia
trial data

SBB ISB Control Total

ni 19 19 15 53
Ti 266 403 485 1154
x Ti ni 14 21 32 22
Ti 2 ni 3724 8548 15682 27954
Si 5600 10485 16499 32584

The formulae in the ANOVA table above can then be used to obtain the between and
within groups sum of squares (Table 1).

Between groups sum of squares = 27954 – (1154*1154)/53 = 2827

Within groups sum of squares = 32584 – 27954 = 4630

These can be used to construct an ANOVA table (Table 2) and calculate the F
statistic. This F statistic can then be used to obtain a P-value. For the current
example, for an F statistic with 2 and 50 degrees of freedom the P-value is < 0.001.
This is highly significant and indicates that there is sufficient evidence to reject the
null hypothesis, that the group means are the same and accept the alternative. We
would conclude that there are systematic differences between the groups.

Table 2: ANOVA table for analgesia trial data

Sum of squares Degrees of Mean F statistic

freedom square (variance ratio)

Between 2827 2 1413.5 15.26

groups
4630 50 92.6
Within groups
Total 7457 52

Comparing different groups

Having decided that there are differences between the groups, it might also be of
interest to test for contrasts between the groups i.e. compare the groups with each
other. One of the advantages of the one-way ANOVA over the t-test is that the
standard error for the difference between pairs of groups is based upon the within
group mean square. As this has more degrees of freedom than a standard error
based upon the two groups alone, the test has greater power to find a difference if
one exists. Many different methods for making post hoc comparisons between
groups have been proposed, all of which are designed to ensure that the overall type
1 error rate stays below 5%. This will be discussed further in a future tutorial, but
briefly, if the numbers in each group are the same then the Student-Newman Keuls
test is recommended, otherwise use Scheffé(Armitage, Berry, & Matthews 2002). If
there is a single control group that you want to test against all other groups (but not

Page 76
the other groups against each other) then Dunnett‟s test can be used. These are all
readily obtained from the major statistical packages.

Presentation of the results of an ANOVA

When presenting the results of an ANOVA it is good practice to report the group
means and the numbers in each group, together with the F statistic, its df and the P-
value associated with this statistic. If any post hoc analyses are carried out, the
procedure used for the pairwise comparisons (e.g. Scheffé) should be stated and the
mean differences between groups together with the associated 95% confidence
intervals for these differences should be presented(Lang and Secic 1997). Thus for
the analgesia trial, where it was of interest to compare the difference between the
two block groups, the results are presented in table 3.

Table 3: Presentation of results of ANOVA

Mean (95% CI) P-value

Study group
ISB (n=19) 14.0 (9.1 to 18.9) < 0.001*
SBB (n=19) 21.2 (16.2 to 26.2)
Control (n=15) 32.3 (28.1 to 36.6)

Difference (SBB – 7.2 (-0.7 to 15.1) 0.079**

ISB)
* ANOVA, F2,50 = 15.26
** Scheffé method used for post hoc comparison

Page 77
Page 78
Sample Size Calculations for Clinical Trials: parallel,
two-group clinical trial with a Normally distributed
outcome variable.
S. A. Julious, J. V Freeman

Introduction

When planning any study, it is important to have a good idea of how many individuals
are needed and this is one of the questions that statisticians get asked most frequently.
The justification for the chosen sample size can range from a formal calculation based
upon a clinical outcome, to what is feasible. Even for the latter, a sample size
justification can be provided as it is possible to determine, for a fixed sample size, what
difference could be detected for a given level of power.

This tutorial describes the simplest case, when calculating the sample size needed for
a two group comparison with a Normally distributed outcome variable, designed to
assess superiority (i.e. whether one group is better than the other), with equal numbers
allocated between groups (i.e. an allocation ratio of 1:1).

General principles

In order to calculate the size of the sample required several quantities need to be
known or estimated a priori, including the number of groups to be compared, the
outcome variable (and its distribution), the anticipated size of the difference between
groups and an estimate of the population variability. In addition the researcher must
decide in advance the maximum acceptable values for the Type I and Type II error
rates (see tutorial 4 (Freeman & Julious 2006b)).

Null and Alternative Hypotheses for a Superiority Trial

A superiority trial is designed to determine whether there is evidence of a statistically

significant difference between treatment groups for the outcome of interest,
compared to the null hypothesis that the treatments are the same. The purpose of
the sample size calculation for this type of study is to provide sufficient evidence to
reject the null hypothesis when in fact some alternative hypothesis is true.

The null (H0) and alternative (H1) hypotheses take the form:

H0: The two treatments are not different with respect to the mean response
( A B ).

H1: The two treatments are different with respect to the mean response
( A B ).

In the definition of the null and alternative hypotheses A and B refer to the mean
response for treatments A and B respectively.

Page 79
Estimate of effect size

In an ideal world the anticipated size of the treatment difference is estimated from pilot
data and this is one of the main reasons for conducting a pilot study. Where no pilot
data exist it may be possible to estimate the effect size from the published literature. In
terms of superiority an estimate of effect size could be based upon an assessment of
what is a clinically meaningful difference. However, it is important that if this is the
case, the difference should be realistic. It is no good setting an effect size for the
sample size calculation that is unrealistically large, as, if the observed difference is
smaller than this the study may well fail to reject the null hypothesis to an appropriate
level of statistical significance.

Estimate of population variance

Once the estimated effect size has been determined we need to have an estimate of
the of the population variability about this effect size, 2 . The smaller the variability the
fewer individuals are required for (all other things being equal)
.
Type I and type II error rates

In any investigation, we are attempting to find out what is true, although in reality the
best we can hope for is to decide what is most likely. However, it is possible that in
making a decision we make an error:

We could reject the null hypothesis (in the case of a superiority trial: that there is
no difference between treatments) even though it is true. This is known as a
Type I error or a false positive.
We could fail to reject the null hypothesis even though it is false - there
genuinely is a difference between treatments. This is known as a Type II error
or false negative error(Freeman & Julious 2006b).

In order to limit the possibility of committing either error, we set the maximum
acceptable level for each a priori. The probability of committing a Type I or false
positive error is given by the greek letter α and we can reduce the risk of committing
this type of error by decreasing the value of α of at which a result is declared
„statistically significant‟. The significance level is usually set at the “magic” 5% or 0.05.
If greater proof is required this could be reduced. For example a significance level of
0.01 would mean we have a Type I error of 1%.

The Type I error is often referred to as a regulatory or society risk as it is upon these
that the cost of this error is incurred if a new regimen, in the case of drug trials, were to
enter the market needlessly. Even outside of drug trials it is a societal risk as a Type I
error may initiate further research in a research area which is actually a dead end.

For a superiority trial there are two chances of rejecting the null hypothesis and thus
making a Type I error. The null hypothesis can be rejected if A B or if A B
by a statistically significant amount. As there are two chances of rejecting the null
hypothesis the statistical test is referred to as a two tailed test and each tail is
allocated an equal amount of the Type I error (=α/2). The sum of these tails adds up
to the overall Type I error rate (α). Thus, for a trial with an overall type I error rate of
5%, the null hypothesis can be rejected if either the test of A B is statistically

Page 80
significant at the 2.5% level or the test of A B is statistically significant at the
2.5% level.

The probability of making a Type II or false negative error is given by the greek letter β.
In general the value of β is set between 0.1 to 0.2. The Type II error is often referred to
as the sponsor‟s risk as it is the sponsor that will incur the cost of this error. The
acceptable value of β is usually greater than that for α as the cost to society from this
error is usually lower.

Often instead of referring to the Type II error, reference is made to the power of a
study. The power is the probability that we will detect a difference of a specified size, if
there is one. That is, it is the probability of not making a Type II error i.e. 1- β. The
power therefore is the probability of rejecting the null hypothesis when it is false.
Although we have said above that the Type II error is usually set between 0.1 and 0.2,
we would recommend that 0.1 as used as the standard level. Often people talk in
terms of power and reducing the power from 0.90 to 0.80 does not seem such a great
step to make but in effect we are doubling the type II error for little actual reduction
(around 25%) in the sample size.

Table 1: Effect on sample size of changing the size of the constituents used in
the sample size calculation

Increase in: Effect on sample

size:
Type I error Decrease
Type II error Decrease
Effect size Decrease
Variance of effect Increase
size

Derivation of the formula for sample size calculation

Figure 1 shows the distribution of the mean difference d, if the null hypothesis is true
(i.e. d=0; the left-hand curve) and if the alternative hypothesis is true (right-hand
curve). We set a critical value (dcrit) for the test such that for any value less than the
critical value we would accept the null hypothesis, whilst for any value greater than
the critical value we would reject the null hypothesis. For a given critical value:

If the alternative hypothesis is true, the dark shaded area shows the
probability of making a type II error, β (i.e. falsely concluding there is no
difference even though there is a difference).
If the null hypothesis is true the lighter shaded area shows the probability of
making a type I error α (i.e. falsely concluding that there is a difference even
though there is no difference).

Page 81
Figure 1: Distribution of the mean difference under the null and alternative
hypotheses

If we assume that both A and B come from populations with the same variance σ 2,
and there are equal numbers in each group (=m) then the standard error of the
difference d is equal to 2 m . If the null hypothesis is true we usually allow for the
possibility of a Type I error in the either direction and so under the null hypothesis that
the mean difference is 0 the critical value is determined by:

2
(1) d crit Z1 /2
m

On the other hand, if the alternative hypothesis is true, the mean difference is Normally
distributed with a mean of d and a standard error of σ√(2/m) and the critical value is
determined by:

2
(2) d crit d Z1
m

Thus

2 2
(3) Z 1 /2 d Z1 ,
m m

and the sample size per treatment group can be estimated by re-arranging (3)

Page 82
 2( Z1 /2 Z1 ) 2 2

(4) m
d2

where 2 is the population variance, d is the effect size of interest and m is the sample
size per group. Z 1 / 2 and Z 1 are Z values obtained from the standard Normal
distribution, for the pre-determined values of α and β. For example, when α=0.05, then
1- α/2=0.975 and the corresponding Z value is 1.96 and when β=0.10 then 1-β=0.9
and the corresponding Z value is 1.282.

The result (4) has a number of advantages not least that the calculations are relatively
simple and easy to do by hand. However, it is worth noting that it is an approximation
and many sample size packages use a slightly different formula based not on the
Normal distribution but on the t distribution. When the sample size is small as may
often be the case with imaging studies you could add a correction factor of Z `21 / 2 4 to
(4) to allow for the approximation (2).
2 2
2 Z1 Z1 /2 Z 12 / 2
(5) m .
d2 4

Table 1 gives sample sizes per group for various standardised differences ( d/ )
using (5).

Table 4. Sample sizes for per group in a parallel group study for different
standardised differences for 90% power and a two sided type I error of 5%
Sample Size
0.05 8407
0.10 2103
0.15 935
0.20 527
0.25 338
0.30 235
0.35 173
0.40 133
0.45 105
0.50 86
0.55 71
0.60 60
0.65 51
0.70 44
0.75 39
0.80 34
0.85 31
0.90 27
0.95 25
1.00 22

For quick calculations the following formula to calculate a sample size with 90%
power and a two-sided 5% type I error rate, can be used

Page 83
 21 2
(6) m .
d2

While for 80% power and a two-sided 5% type I error rate we can use

16 2
(7) m ,
d2

Worked example 1:

Researchers wished to assess which of two methods of managing post-operative

pain following shoulder surgery was the most effective in reducing the post-operative
experience for patients. The main outcome measure was the amount of patient-
administered morphine used in the first 24 hours post-op. The effect size of interest
was 7mg morphine and the population standard deviation was estimated as 14mg.
The researchers wanted to have 90% power to detect a difference with a two sided
significance of 0.05. The calculations are as follows:
0.1, Z1 Z 1 0.1 Z 0.9 1.282
0.05 Z1 /2 Z1 0.025 Z 0.975 1.96
2 2
14 196
2 2
d 7 49
2
2 1.282 1.96 196 1.96 2
m 84.08 0.96 85.04 86
49 4
Thus, the number of patients needed per group was 86 and the total number needed
in both groups was 2x86 = 172 patients.

Alternatively we can calculate d/ =7/14=0.5 and from Table 1 we also obtain a

sample size estimate of 172.

Accounting for drop-outs:

Described above is a method for calculating the required sample size of evaluable
subjects (the minimum number required for the statistical analysis). However, often this
is only the first step in estimating the size of a study. Often we can anticipated drop-
outs in a study and this needs to be accounted for in the final sample size of patients to
be recruited.

To account for drop outs we need to anticipated likely dropout rate. Suppost we
anticipate that a proportion p subject to drop out. The recruited sample size (to ensure
m evaluable subjects) would then be

m
(8) mRe cruited
1 p

Page 84
Worked example 2:

Revisiting the worked example. suppose the researchers anticipated a drop-out rate of
about 20%, the number needed to be recruited in total would need to be 172/(1-
0.2)=215.

Sample size based on feasibility

We have just shown how to calculate a sample size for a given effect size. However,
it can sometimes be the case that the sample size is fixed by practical considerations
or feasibility, even before a study takes place. When this is the case, the trial is not
powered to detect any pre-specified effect, but a power analysis can still be used to
determine the difference that could be detected for this fixed sample size. In this
case, as we now have m, 2, β and α, we can re-arrange (1) to calculate the
difference (d) that could be detected for this fixed sample size, m, where m is the
number per group:

2
(9) d ( Z1 /2 Z1 )
m

If sample size is based primarily on feasibility then it should be clearly highlighted as

such in any protocol.

Summary

This article describes how to calculate the sample size for a Normally distributed
outcome for a two group comparison designed to assess superiority as. A sample size
justification should be provided for all clinical studies even where the sample size is
based on feasibility

Page 85
Page 86
Diagnostic Tests
Michael J Campbell, Jenny V Freeman

In this tutorial we will examine how to evaluate a diagnostic test. Initially we will
consider the case when there is a binary measure (two categories: disease present /
disease absent). We will then look at how define a suitable cut-off for an ordinal or
continuous measurement scale and we will finish with a short discussion contrasting
diagnostic tests with screening tests.

When evaluating any diagnostic test one should have a definitive method for deciding
whether the disease is present in order to see how well the test performs. For
example, to diagnose a cancer one could take a biopsy, to diagnose depression one
could ask a psychiatrist to interview a patient, and to diagnose a walking problem
one could video a patient and have it viewed by an expert. This is sometimes called
the „gold standard‟. Often the gold standard test is expensive and difficult to
administer and thus a test is required that is cheaper and easier to use.

Binary situation

Let us consider first the simple binary situation in which both the gold standard and
the diagnostic test have either a positive or negative outcome (disease is present or
absent). The situation is best summarised by the 2x2 table below (Table 1). In writing
this table always put the gold standard on the top, and the results of the test on the
side:

Table 1. Standard table for diagnostic tests

Gold Standard
Positive Negative
Diagnostic Test Positive a b a+b
Negative c d c+d
Total a+c b+d n

The numbers „a‟ and „d‟ are the numbers of true positives and true negatives
respectively. The number „b‟ is the number of false positives, because although the
test is positive the patients don‟t have the disease, and similarly „c‟ is the number of
false negatives. The prevalence of the disease is the proportion of people diagnosed
by the gold standard and is given by (a+c)/n, although this is often expressed as a
percentage.

In order to assess how good the test is we can calculate the sensitivity and
specificity, and the positive and negative predictive values. The sensitivity of the test
is the proportion of people with the disease who are correctly identified as having the
disease. This is given by a/(a+c) and is usually presented as a percentage. Suppose
a test is 100% sensitive. Then the number of false negatives is zero and we would
expect Table 2.

Page 87
Table 2 Results of a diagnostic test with 100% sensitivity

Gold Standard
Positive Negative
Diagnostic Test Positive a b a+b
Negative 0 d d
Total a b+d n

From Table 2 we can see that if a patient has a negative test result we can be certain
that the patient does not have the disease. Sackett et al(2007) refer to this as SnNout
i.e. for a test with a high sensitivity (Sn), a Negative result rules out the disease.

The specificity of a test is the proportion of people without the disease who are
correctly identified as not having the disease. This is given by d/(b+d) and as with
sensitivity is usually presented as a percentage. Now suppose a test is 100%
specific. Then the number of false positives is zero and we would expect table 3

Table 3 Results of a diagnostic test with 100% specificity

Gold Standard
Positive Negative
Diagnostic Test Positive a 0 a
Negative c d c+d
Total a+c d n

From Table 3 we can see that if a patient has a positive test we can be certain the
patient has the disease. Sackett et al (x) refer to this as SpPin., i.e. for a test with a
high specificity (Sp), a Positive test rules in the disease.

Box 1: Useful Mnemonic

SeNsitivity=1-proportion false Negatives (n in each

side)
SPecificity=1-proportion false Positives (p in each
side)

What patients really want to know, however, is „if I have a positive test, what are the
chances I have the disease?‟ This is given by the positive predictive value (PPV)
which is a/(a+b). One way of looking at the test is that before the test the chances of
having the disease was (a+c)/n. After the test they are either a/(a+b) or c/(c+d)
depending on whether the result was positive or negative.

The negative predictive value is the proportion of those whose test result is negative
who do not have the disease and is given by d/(c+d).

It should be noted that whilst sensitivity and specificity are independent of

prevalence, positive and negative predictive values are not. Sensitivity and
specificity are characteristics of the test and will be valid for different populations with
different prevalences. Thus we could use them in populations with high prevalence
such as elderly people as well as for low prevalence such as for young people.

Page 88
However, the PPV is a characteristic of the population and so will vary depending on
the prevalence.

To show this suppose that in a different population, the prevalence of the disease is
double that of the current population (assume the prevalence is low, so that a and c
are much smaller than b and d and thus the results for those without the disease are
much the same as the earlier table). The situation is given in Table 4

Table 4 Standard situation but with a doubling of the prevalence

Gold Standard
Positive Negative
Diagnostic Test Positive 2a b 2a+b
Negative 2c d 2c+d
Total 2(a+c) b+d n'

The sensitivity is now 2a/(2a+2c)=a/(a+c) as before. The specificity is unchanged.

However the positive predictive value is given by 2a/(2a+b) which is greater than the
earlier value of a/(a+b).

Likelihood ratio

It is common to prefer a single summary measure, and for a diagnostic test this is
given by the likelihood ratio for a positive test (LR(+) as defined below:

LR+ = Probability of positive test given the disease = Sensitivity = a (b+d)

Probability of positive test without disease 1-Specificity b (a+c)

One reason why this is useful is that it can be used to calculate the odds of having
the disease given a positive result. The odds of an event are defined as the ratio of
the probability of the event occurring to the probability of the event not occurring i.e.
p/(1-p) where p is the probability of the event. Before the test is conducted the
probability of having the disease is just the prevalence, and the odds are simply
[(a+c)/n]/[b+d)/n]= (a+c)/(b+d). The odds of having the disease after a positive test
are given by

Odds of disease after positive test=odds of disease before test x LR(+)= a/b

We can also get the odds of disease after a positive test directly from the PPV since
the odds of disease after a positive test is PPV/(1-PPV).

Example:

A recent study by Kroenke et al (Kroenke et al. 2007) surveyed 965 people attending
primary care centres in the US. They were interested in whether a family practitioner
could diagnose Generalized Anxiety Disorder (GAD) by asking two simple questions
(the GAD2 questionnaire): „Over the last two weeks, how often have you been
bothered by the following problems? 1) feeling nervous, anxious or on edge 2) Not
able to stop or control worrying. The patients answered each question from „not at
all‟, „several days‟, „more than half‟ and „nearly every day‟, scoring 0,1,2 or 3
respectively. The scores for the two questions were summed and a score of over 3

Page 89
was considered positive. Two mental health professionals then held structured
psychiatric interviews with the subject over the telephone to diagnose GAD. The
professionals were ignorant of the result of the GAD2 questionnaire. The results are
given in table 2:

Table 4.2 Results from Kroenke et al(Kroenke, Spitzer, Williams, Monahan, & Löwe
2007)

Diagnosis by mental health

worker
Positive Negative
GAD2 ≥3 63 152 215
<3 10 740 750
Total 73 892 965

The prevalence of the disease is given by (a+c)/n=73/965=0.076=7.6%.

The sensitivity of the test is given by a/(a+c)= 63/73=0.86=86%.

The specificity of the test is given by d/(b+d) = 740/892=0.83=83%

The positive predictive value (PPV) which is a/(a+b) = 63/215==0.29=29%.

The negative predictive value is d/(c+d)= 740/750=0.987=98.7%.

Thus before the test the chances of having GAD were 7.6%. After the test they are
either 29% or 1.3% (i.e. 100*(1-0.987) depending on the result. Note that even with a
positive test the chances of having GAD are still less than 1/3.

For the GAD example we find that LR(+)= 0.86/(1-0.83)=5.06 and the odds=0.29/(1-
0.29)=0.41.

ROC Curves

For a diagnostic test that produces results on a continuous or ordinal measurement

scale a convenient cut-off level needs to be selected to calculate the sensitivity and
specificity. For example the GAD2 questionnaire has possible values from 0 to 6.
Why should one choose the value of 3 as the cut-off? For a cut off of 2 the sensitivity
is 0.95, the specificity is 0.64 and the LR(+) is 2.6(Kroenke, Spitzer, Williams,
Monahan, & Löwe 2007). One might argue that since a cut-off of 3 has a better
LR(+) then one should use it. However, a cut-off of 2 gives a higher sensitivity, which
might be important. It should be noted that a sensitivity of 100% is always achievable
by stating that everyone has the disease, but this is at the expense of a poor
specificity (similarly a 100% specificity can be achieved by stating no-one has the
disease. If the prevalence is low, this tactic will have a high accuracy, i.e. it will be
right most of the time, but sadly wrong for the important cases). A discussion of the
different scenarios for preferring a high specificity or sensitivity is given in the next
section.

A simple graphical device for displaying the trade-offs between sensitivity and
specificity for tests on a continuous or ordinal scale is a receiver operating

Page 90
characteristics (ROC) curve (the unusual name originates from electrical
engineering). This is a plot of sensitivity versus one minus specificity for different cut-
off values of the diagnostic test. ROC curves for two theoretical tests are shown in
Fig 1, together with the line of equality which is what we would expect if a test had no
power to detect disease. A perfect diagnostic test would be one with no false
negatives (i.e. sensitivity or 1) or false positives (i.e. specificity of 1) and would be
represented by a line starting at the origin, travelling vertically up the Y-axis to a
sensitivity of 1 and then horizontally across to a false positive rate of 1. Any
diagnostic test that was reasonable would produce a ROC curve in the upper left-
hand triangle of figure 1. The selection of the optimal cut-off will depend upon the
relative medical consequences and costs of false positive and false negative errors.

ROC curves are particularly useful for comparing different diagnostic tests and when
more than one test is available they can be compared by plotting both on the same
plot. A test for which the plot is consistently nearer the left hand side and the top is to
be preferred. In addition the area under the curve (AUC) for each plot can be
calculated. For the perfect test outlined above the AUC is 1 and represents the total
area of the panel (i.e. 1x1). For the two curves displayed it is obvious that the best
test is the one with the line represented by the dashed line on the left of the Figure.
This has an AUC value of 0.95 compared to the other much poorer fitting line which
as an AUC value of 0.59.

Fig 1: Example ROC curves showing also the line of equality

1.0

0.8

0.6
Sensitivity

0.4

0.2
Poor diagnostic test; AUC=0.59
Good diagnostic test; AUC=0.95

0.0
0.0 0.2 0.4 0.6 0.8 1.0

1-Specificity

Distinction between diagnosis and screening.

It is important to understand the difference between diagnosing a disease and

screening for it. In the former case there are usually some symptoms, and so there

Page 91
may already be a suspicion that something is wrong. If a test is positive some action
will be taken. In the latter case there are usually no symptoms and so if the test is
negative the person will have no further tests. Recalling Sackett‟s mnemonics SpPin
and SnNout, for diagnosis we want a positive test to rule people in, so we want a
high specificity. For screening we want a negative test to rule people out so we want
a high sensitivity. Thus mass mammography will have a fairly low threshold of
suspicion, to ensure a high sensitivity and reduce the chances of missing someone
with breast cancer. The subsequent biopsy of positive results will have a high
specificity to ensure that if, say, mastectomy is to be considered, the doctor is almost
certain that the patient has breast cancer.

Summary

This tutorial has summarised the methods used for examining the suitability of a
particular test for diagnosing disease. In addition it has highlighted the difference
between diagnostic and screening tests. In reality the same methods are used to
evaluate both diagnostic and screening tests, the important difference being the
emphasis that is placed on the sensitivity and specificity. Further details are given in
Campbell, Machin and Walters (Campbell et al. 2007) Chapter 4.

Page 92
Correlation Coefficient
Dr Jenny Freeman, Dr Tracey Young

Many statistical analyses can be undertaken to examine the relationship between two
continuous variables within a group of subjects. Two of the main purposes of such
analyses are:
To assess whether the two variables are associated. There is no distinction
between the two variables and no causation is implied, simply association.
To enable the value of one variable to be predicted from any known value of
the other variable. One variable is regarded as a response to the other
predictor (explanatory) variable and the value of the predictor variable is used
to predict what the response would be.
For the first of these, the statistical method for assessing the association between
two continuous variables is known as correlation, whilst the technique for the second,
prediction of one continuous variable from another is known as regression.
Correlation and regression are often presented together and it is easy to get the
impression that they are inseparable. In fact, they have distinct purposes and it is
relatively rare that one is genuinely interested in performing both analyses on the
same set of data. However, when preparing to analyse data using either technique it
is always important to construct a scatter plot of the values of the two variables
against each other. By drawing a scatter plot it is possible to see whether or not there
is any visual evidence of a straight line or linear association between the two
variables.

This tutorial will deal with correlation, and regression will be the subject of a later
tutorial.

Correlation

The correlation coefficient is a measure of the degree of linear association between

two continuous variables i.e. when plotted together, how close to a straight line is the
scatter of points. No assumptions are made about whether the relationship between
the two variables is causal i.e. whether one variable is influencing the value of the
other variable; correlation simply measures the degree to which the two vary
together. A positive correlation indicates that as the values of one variable increase
the values of the other variable increase, whereas a negative correlation indicates
that as the values of one variable increase the values of the other variable decrease.
The standard method (often ascribed to Pearson) leads to a statistic called r,
Pearson's correlation coefficient. In essence r is a measure of the scatter of the
points around an underlying linear trend: the closer the spread of points to a straight
line the higher the value of the correlation coefficient; the greater the spread of points
the smaller the correlation coefficient. Given a set of n pairs of observations (x1, y1),
(x2, y2),..(xn, yn) the formula for the Pearson correlation coefficient r is given by:

Page 93
Certain assumptions need to be met for a correlation coefficient to be valid as
outlined in Box 1. Both x and y must both be continuous random variables, (and
Normally distributed if the hypothesis test is to be valid).

Pearson's correlation coefficient r can only take values between -1 and +1; a value of
+1 indicates perfect positive association (Figure 1), a value of -1 indicates perfect
negative association (Figure 2), and a value of 0 indicates no linear association
(Figure 3).

Figure 1: perfect negative correlation (r =-1)

Figure 2: perfect negative correlation (r =-1)

Page 94
Figure 3: no linear association (r =0)

The easiest way to check whether it is valid to calculate a correlation coefficient is to

examine the scatterplot of the data. This plot should be produced as a matter of
routine when correlation coefficients are calculated, as it will give a good indication of
whether the relationship between the two variables is roughly linear and thus whether
it is appropriate to calculate a correlation coefficient all. In addition, as the correlation
coefficient is highly sensitive to a few abnormal values, a scatterplot will show
whether this is the case, as illustrated below in Figures 4 & 5:

Figure 4: The correlation for this plot is 0.8. It is heavily influenced by the
extreme cluster of four points away from the main body

Figure 5: The correlation for this plot is close to 0. Whilst it is clear that the
relationship is not linear and so a correlation is not appropriate, it is also clear
that there is a strong n shaped relationship between these two variables.

Page 95
Example

Consider the heights and weights of 10 elderly men:

(173, 65), (165, 57), (173, 77), (183, 89), (178, 93), (188, 73), (180, 83), (183, 86),
(163, 70), (178, 83)

Plotting these data indicates that, unsurprisingly, there is a positive linear relationship
between height and weight. The shorter a person is the lower their weight and
conversely, the taller a person is the greater their weight. In order to examine
whether there is an association between these two variables, the correlation
coefficient can be calculated. In calculating the correlation coefficient, no
assumptions are made about whether the relationship is causal i.e. whether one
variable is influencing the value of the other variable.

Figure 7: Plot of weight against height for ten elderly men

Subject x x x ( x x) 2 y y y ( y y) 2 (x x )( y y )
1 173 -3.5 12.25 65 -12.5 156.25 43.75
2 165 -11.5 132.25 57 -20.5 420.25 235.75
3 174 -2.5 6.25 77 -0.5 0.25 1.25
4 183 6.5 42.25 89 11.5 132.25 74.75
5 178 1.5 2.25 93 15.5 240.25 23.25
6 188 11.5 132.25 73 -4.5 20.25 -51.75
7 180 3.5 12.25 83 5.5 30.25 19.25
8 182 5.5 30.25 86 8.5 72.25 46.75
9 163 -13.5 182.25 70 -7.5 56.25 101.25
10 179 2.5 6.25 82 4.5 20.25 11.25
Total 1765 0.0 558.50 775 0.0 1148.50 505.50

x =1765 / 10 = 176.5 cm
y =775 / 10 = 77.5 kg
r = 505.50 / √(558.50*1148.50) = 0.63

Page 96
Thus the Pearson correlation coefficient for these data is 0.63, indicating that there is
a positive association between height and weight for these 10 men. When calculating
the correlation coefficient it is assumed that at least one of the variables is Normally
distributed. If the data do not have a Normal distribution, a non-parametric correlation
coefficient, Spearman's rho (rs), can be calculated. This is calculated in the same way
as the Pearson correlation coefficient, except that the data are ordered by size and
given ranks (from 1 to n, where n represents the total sample size) and the
correlation is calculated using the ranks rather than the actual values. For the data
above the Spearman correlation coefficient is 0.59.

Subject Rank x x (x x) 2 Rank (y) y y (y y) 2 (x x )( y y)

(x)
1 3 (173) -2.5 6.25 2 (65) -3.5 12.25 8.75
2 2 (165) -3.5 12.25 1 (57) -4.5 20.25 15.75
3 4 (174) -1.5 2.25 5 (77) -0.5 0.25 0.75
4 9 (183) 3.5 12.25 9 (89) 3.5 12.25 12.25
5 5 (178) -0.5 0.25 10 (93) 4.5 20.25 -2.25
6 10 4.5 20.25 4 (73) -1.5 2.25 -6.75
7 (188) 1.5 2.25 7 (83) 1.5 2.25 2.25
8 7 (180) 2.5 6.25 8 (86) 2.5 6.25 6.25
9 8 (182) -4.5 20.25 3 (70) -2.5 6.25 11.25
10 1 (163) -0.5 0.25 6 (82) 0.5 0.25 0.25
6 (179)
Total 55 0.0 82.50 55 0.0 82.50 48.5

x (ranks) = 55 / 10 = 5.5
y (ranks) = 55 / 10 = 5.5
rs = 48.5 / √(82.5*82.5) = 0.59

The square of the correlation coefficient gives the proportion of the variance of one
variable explained by the other. For the example above, the square of the correlation
coefficient is 0.398 indicating that about 39.8% of the variance of one variable is
explained by the other.

Hypothesis testing

The null hypothesis is that the correlation coefficient is zero. However, its significance
level is influenced by the number of observations and so it is worth being cautious
when comparing correlations based on different sized samples. Even a very small
correlation can be statistically significant if the number of observations is large. For
example, with 10 observations, a correlation of 0.63 is significant at the 5% level,
whereas with 150 observations, a correlation of 0.16 is significant at the 5% level.
Figure 7a& b illustrate this:

Page 97
 Correlation = 0.16, P=0.04 (n=150)
Correlation = 0.63, p=0.048 (n=10)

The statistical test is based on the test statistic t = r / se(r) which under the null
hypothesis follows a Students‟ t distribution on n-2 degrees of freedom and the
confidence interval is given by:

1 r2
The standard error of r =
n 2

For the Pearson correlation coefficient above the standard error is 0.27, the t statistic
is 2.30 and the P-value is 0.05.

Box 1: The assumptions underlying the validity of the hypothesis test

associated with the correlation coefficient

1. The two variables are observed on a random sample of individuals.

2. The data for at least one of the variables should have a Normal
distribution in the population.
3. For the calculation of a valid confidence interval for the correlation
coefficient both variables should have a Normal distribution.

When not to use a correlation coefficient

Whilst the correlation coefficient is a useful measure for summarising how two
continuous variable are related, there are certain situations when it should not be
calculated, as has already been alluded to above. As it measures the linear
association between two variables, it should not be used when the relationship is
non-linear. Where outliers are present in the data, care should be taken when
interpreting its value. It should not be used when the values of one of the variables
are fixed in advance, for example when measuring the responses to different doses
of a drug. Causation should not be inferred from a correlation coefficient. There are
many other criteria that need to be satisfied before causation can be concluded.
Finally, just because two variables are correlated at a particular range of values, it
should not be assumed that the same relationship holds for a different range.

Page 98
Summary

This tutorial has outlined how to construct the correlation coefficient between two
continuous variables. However, correlation simply quantifies the degree of linear
association (or not) between two variables. It is often more useful to describe the
relationship between the two variables, or even predict a value of one variable for a
given value of the other and this is done using regression. If it is sensible to assume
that one variable may be causing a response in the other then regression analysis
should be used. If on the other hand, there is doubt as to which variable is the causal
one, it would be most sensible to use correlation to describe the relationship.
Regression analysis will be covered in a subsequent tutorial.

Page 99
Page 100
Simple Linear Regression
In the previous tutorial we looked at using correlation to assess the strength of the
linear relationship between two continuous variables(Freeman and Young 2009). The
correlation coefficient simply measures the strength of the linear association as a
single number. No distinction is drawn between the two variables and no causation is
implied. However, it is often of interest to quantify the relationship between two
continuous variables, and given the value of one variable for an individual, to predict
the value of the other variable. This is achieved using the technique known as simple
linear regression. One variable is regarded as a response to the other predictor
(explanatory) variable and the value of the predictor variable is used to predict what
the response would be.

Scatter plots

As stated in the previous tutorial, when undertaking either a correlation or simple

linear regression analysis it is important to construct a scatter plot of the data. The
values of one variable are plotted on the horizontal axis (known as the x-axis) and
the values of another are plotted on the vertical axis (y-axis). By drawing a scatter
plot it is possible to see whether or not there is any visual evidence of a straight line
or linear association between the two variables. It is possible for there to be a
relationship between two variables but for that relationship to not be linear. In this
case correlation or simple linear regression analysis may not be the most appropriate
methods to use. In addition a scatterplot provides a good way of examining the data
and checking for outliers or odd values.

If it is known (or suspected) that the value of one variable (known as the independent
variable) influences the value of the other variable (known as the dependent
variable), it is usual to plot the independent variable on the horizontal axis and the
dependent variable on the vertical axis. In the case of height and weight, as height
determines weight, to an extent, and not the other way around, a scatterplot of weight
against height would be plotted with height on the horizontal axis and weight on the
vertical axis.

Simple linear regression

In the technique of simple linear regression a straight-line equation is used to model

the relationship between the predictor variable and the response variable. The
equation of the regression line is given by:

y= a +bx

Where :
x = independent / predictor / explanatory variable: variable that is used to
predict the values of the response variable. This is plotted on the
horizontal axis of a scatter plot
y = dependent / response / outcome variable: variable being predicted by the
model. This is plotted on the vertical axis of a scatterplot
a = intercept. This is the point at which the regression line crosses the
vertical (Y) axis. Strictly speaking this gives the value of the Y variable
(dependent variable) when the X variable (independent variable) is zero.

Page 101
 b = regression coefficient. It is also known as the slope and it shows the
average change in the Y variable (outcome) for a unit change in the X
variable (predictor/explanatory variable)

a and b are calculated as follows:

n
xi x yi y
i 1
b n
2
xi x
i 1

a y bx

Example: Simple linear regression of weight against height for ten elderly men

The figure below shows the height and weight values for ten elderly men. The data
are given in Table 1

Figure 1: Scatter plot of weight against height together with the regression line

y variable: dependent or response

variable, plotted on the vertical axis
95

90

85

80
Weight (kg)

75

70

65

The regression line: its slope gives the

60 average change in the y variable (in this
case: weight) for a change in the x variable
(height) of one unit
55
160 165 170 175 180 185 190

Height (cm)
x variable: independent or explanatory
variable, plotted on the horizontal axis

Page 102
Table 1: Calculation of regression equation for regression of weight on height
of 10 elderly men

Subject x (x- x ) (x- x )2 y (y- y ) (y- y )2 (x- x )(y- y )

1 173 -3.5 12.25 65 -12.5 156.25 43.75
2 165 -11.5 132.25 57 -20.5 420.25 235.75
3 174 -2.5 6.25 77 -0.5 0.25 1.25
4 183 6.5 42.25 89 11.5 132.25 74.75
5 178 1.5 2.25 93 15.5 240.25 23.25
6 188 11.5 132.25 73 -4.5 20.25 -51.75
7 180 3.5 12.25 83 5.5 30.25 19.25
8 182 5.5 30.25 86 8.5 72.25 46.75
9 163 -13.5 182.25 70 -7.5 56.25 101.25
10 179 2.5 6.25 82 4.5 20.25 11.25
Total 1765 0 558.5 775 0 1148.5 505.5

x =1765/10 = 176.5cm

y =775/10 = 77.5 kg

b = 505.5 / 558.5 = 0.9051

a = 77.5 – 0.905103*176.5 = -82.25

Thus the regression equation for these data is:

weight = -82.25 +0.9051 * height

From this it can be seen that the slope coefficient was 0.9051, indicating that for
every 1cm increase in height there was an increase in weight of 0.9051 kg. Note that
the value of the intercept is -82.25. Thus when height is zero weight is -
82.25kg. Clearly this is nonsense and illustrates an important principle for regression
analyses: they should never be used to predict values outside of the range of
observations. However, within the range of the data the regression equation can be
used to predict the values of the y variable for particular values of the x variable. For
example the estimated weight for an elderly man who was 180cm tall is calculated as
follows:

weight = -82.25 +0.9051 * 180 = 80.67kg

Assumptions and model fit

Three important assumptions underlie a simple linear regression analysis as outlined

in Box 1 and as with any statistical analysis it is important to check that they are valid
and that the model fits the data adequately. The first assumption can be checked by
constructing a scatter plot of the data (Figure 1).

Page 103
Assumptions:

1. The relationship between the two variables should be linear

2. The value of the response variable, y, should have a Normal distribution
for each value of the explanatory variable x
3. The variance (or standard deviation) of y should be the same at each
value of x i.e. there should be no evidence that as the value of y changes,
the spread of the x values changes

The final two assumptions can be checked by examining the residuals from the fitted
model. Each y observation has a residual associated with it; this is the difference
between the actual observed y value (yobs) and the y value predicted by the model
(known as the fitted value (yfit)) (see table 2). In Figure 1 for each point the residual is
given by the vertical distance between that point and the fitted regression line. For
example, for the first observation in Table 1, the actual weight is 65kg and the
predicted weight is 74.33kg, thus the residual is given by 65 – 74.33= -9.33kg.

Table 1: Calculation of residuals from the fitted model

Actual height Actual weight Predicted value = Residual

(m) (kg) -82.25+ height*0.9051 (yobs -yfit)
173 65 74.33 -9.33
165 57 67.09 -10.09
174 77 75.24 1.76
183 89 83.38 5.62
178 93 78.86 14.14
188 73 87.91 -14.91
180 83 80.67 2.33
182 86 82.48 3.52
163 70 65.28 4.72
179 83 79.76 2.24

In order for assumption 2 to be valid the residuals should be Normally distributed.

This is most easily checked by constructing a histogram of the residuals to check that
this is approximately Normal (Figure 2). With only 10 individuals it is difficult to
definitively conclude that the residuals are Normally distributed, but given that they
are spread out around a central peak it would appear to be reasonable to accept this
assumption as being valid. In order to check assumption 3 it is necessary to do a
scatter plot of the residuals against the predicted values. This should show a good
spread with no obvious patterns (i.e. it looks random) as in Figure 3 below.

Page 104
Figure 2: Histogram of the residuals from the fitted model

4
Frequency

0
-15.00 -10.00 -5.00 0.00 5.00 10.00 15.00
Residual for Weight

Figure 3: Plot of the residuals from the model against the predicted values

20

15
Residual values (yobs - yfit)

10

-5

-10

-15

-20
60 65 70 75 80 85 90

Predicted value for weight (kg)

R2

The value of r2 is often quoted in published articles and indicates the proportion
(sometimes expressed as a percentage) of the total variability of the outcome
variable that is explained by the regression model fitted. A well fitting model will have

Page 105
a high r2 and a badly fitting model will have a low value of R 2. It is calculated as
follows:

n 2

xi x yi y
2 i 1
r = n n
2 2
xi x yi y
i 1 i 1

(Note that this is also the square of the correlation coefficient:

n
xi x yi y
i 1
)
n n
2 2
xi x yi y
i 1 i 1

For the current example the value of r2 is 0.398. Thus 39.8% of the total variability in
weight for the ten men is explained by their heights.

More than one explanatory variable

Simple linear regression as described above involves the investigation of the effect of
a single explanatory variable on the outcome of interest. However, there is usually
more than one possible explanatory variable influencing the values of the outcome
variable and the method of regression can be extended to investigate the influence of
more than one explanatory variable on the outcome of interest. In this case it is
referred to as multiple regression, and the influence of several explanatory variables
can be investigated simultaneously. This is beyond the scope of the current tutorial
and will be covered in a subsequent tutorial.

Summary: Regression or Correlation?

Regression and correlation are related methods (note that the r2 coefficient is simply
the square of the correlation coefficient!). As they are often presented together it is
easy to get the impression that they are inseparable. In fact, they have distinct
purposes and it is relatively rare that one is genuinely interested in performing both
analyses on the same set of data. Regression is more informative than correlation.
Correlation simply quantifies the degree of linear association (or not) between two
variables. However, it is often more useful to describe the relationship between the
two variables, or even predict a value of one variable for a given value of the other
and this is done using regression. If it is sensible to assume that one variable may be
causing a response in the other then regression analysis should be used.

Page 106
Risk
Jenny Freeman, Dawn Teare

The analysis of data related to risk is important to many fields in medicine,

particularly when explaining different treatment options to patients. There are various
ways in which risk can be measured and this paper will explain some of the more
common measures used including absolute risk, relative risk, odds, odds ratio and
number needed to treat.

Risk data

It is often of interest to know about the risks associated with particular events or
exposures, for example the risk of developing lung cancer for smokers. At the most
basic level risk data are often divided into categories depending on whether
individuals are exposed to the hazard of interest or not, and whether they experience
the event of interest or not. Data such as these can be organised as follows:

Table 1: 2x 2 table illustrating the calculation of risk

Exposure: Total
Yes No
Event:
Yes a b a+b
No c d c+d
a+c b+d n

where:
a= number of individuals who are exposed and have the event of interest
b= number of individuals not exposed who have the event of interest
c= number of individuals exposed who do not have the event of interest
d= number of individuals not exposed who do not have the event

Risk/Absolute risk

The simplest measure of risk is the absolute risk of an event occurring. This is
sometimes simply referred to as the risk and is the number of individuals in the
population under study who experience the event of interest within a defined period
of time divided by the total number of individuals in the group at the start of the time
period.

number who have the event of interest

Absolute risk of an event
number in the group at the start of the follow - up period

For the data in Table 1:

– Absolute risk of event for the exposed group = a/(a+c)
– Absolute risk of event for the unexposed group = b/(b+d)

The absolute risk is a measure of how likely an event is to occur and is a probability.
All probabilities range between 0 and 1, a value of 1 denotes an event that is certain

Page 107
to happen and 0 denotes an event that is never going to happen. It is only possible to
obtain the risk of an event occurring with data that are longitudinal in nature as in
order to calculate the risk of an event in a given period of time it is necessary to know
the total number who were at risk at the start of the time period. Thus it is not
possible to compute a risk for data collected as part of a case-control study (a case-
control study is one in which a group of subjects (cases) with the disease or condition
of interest are compared to a group of subjects (controls) without the disease).

Example 1:

A recent study looked at the effects of the introduction of laparoscopic bariatric

surgery in England. The authors examined the 28 day readmission rates by type of
procedure. Of the 3191 patients who underwent gastric bypass 308 were readmitted
within 28 days, whereas of the 3649 patients who underwent gastric banding 232
were readmitted within 28 days(Burns et al. 2010). These results are reported in the
following 2 x 2 table. The columns represent the type of procedure and the rows
represent readmission within 28 days.

Table 2: 2x 2 table illustrating the calculation of risk

Exposure: Total
Gastric Gastric
bypass banding
Event:
Readmitted within 28 days 308 232 540
Not readmitted 2883 3417 6300
3191 3649 6840

For these data:

– The absolute risk of readmission (within 28 days) for those patients who
underwent a gastric bypass is 308 / 3191 = 0.097
– The absolute risk of readmission (within 28 days) for those patients who
underwent gastric banding is 232 / 3649 = 0.064

Absolute risk difference

This is the absolute additional risk of an event due to a particular exposure. It is

calculated as the risk in the exposed group minus the risk in the unexposed group
(ignoring the sign).

Absolute Risk Difference = |risk in the exposed – risk in the unexposed|5

If the risk is harmful, so that the risk is increased by the exposure this difference is
called the absolute risk excess (ARE) (for example the absolute risk excess for
gastric bypass compared to gastric banding is |0.097 – 0.063| = 0.033) and it
represents the absolute increase in risk for those exposed compared to the
unexposed. If the risk is decreased by the exposure (for example using sunscreen to
reduce the risk of melanoma) then this difference is called the absolute risk
reduction (ARR). A recent randomised controlled trial looking at the risk of
5
The symbols | | are mathematical notation for the modulus and indicate that for any quantity within these only
the absolute value is to be used, the sign is to be ignored. Thus the modulus of -1 (written as |-1|) is 1

Page 108
secondary lymphoedema following treatment for breast cancer compared a group
who had early physiotherapy and education with a control group who had education
alone(Lacomba et al. 2010). At the end of a year‟s follow-up the two groups were
compared for the occurrence of lymphoedema. The data are shown in Table 3:

Table 3: 2x 2 table illustrating the calculation of risk

Exposure: Total
Physiotherapy Control
(n=59) (n=57)
Event: 18
Lymphoedema 4 14
No lymphoedema 55 43 98

The risk of lymphoedema for the physiotherapy group was 4/59 = 0.068 and the risk
of lymphoedema for the control group was 14/57=0.246. Thus the absolute risk
reduction was 0.068-0.246 =| -0.178|=0.178.

Relative risk

The relative risk of a particular event for a given exposure is the ratio of the risk of the
event occurring in the exposed group divided by the risk of the event occurring in the
unexposed group. Using the terminology in Table 1:
a /(a c) a (b d )
Relative risk of an event
b /(b d ) b(a c)
For the data in table 2 the relative risk of readmission for those patients undergoing a
gastric bypass compared to those undergoing gastric banding is 0.097 / 0.064 = 1.52.
Thus patients who undergo gastric bypass are 1.52 times more likely to be
readmitted to hospital within 28 days then patients who undergo gastric banding.

RRR / Relative risk reduction

In clinical trials when looking at the benefits of one treatment compared to another,
the relative risk reduction can also be calculated. This is the extent to which a
treatment reduces a risk in comparison to a group not receiving the treatment of
interest. In this context the risk is an adverse outcome or event. It is calculated as
follows:

risk in the control group - risk in the treated group

Relative risk reduction
risk in the control group
risk in the control group risk in the treated group
-
risk in the control group risk in the control group
1 relative risk

As well as being expressed as a proportion it can also be expressed as a percentage

i.e. 100*(1-RR). In the trial of physiotherapy for the prevention of lymphoedema the
relative risk reduction is 1-(0.068/0.246) = 0.72, or expressed as a percentage 72%.
Thus the relative risk reduction of using physiotherapy to prevent lymphoedema is
72%.

Page 109
Odds

The odds of an event occurring is the ratio of the probability of the event occurring to
the probability of the event not occurring. Using the terminology in Table 1:
Odds of an event given exposure = a/c
Odds of an event given not exposed = b/d

Example

For the data in Table 2 the odds of readmission to hospital within 28 days for patients
undergoing gastric bypass was 308/2883 = 0.107 and the odds of readmission to
hospital within 28 days for patients undergoing gastric banding was 232/3417 = 0.068

Odds ratio

The odds ratio is the ratio of the odds of an event in the exposed group compared to
the unexposed group. Using the terminology of Table 1:

a/c ad
Odds ratio of an event (for exposed compared to not exposed)
b/d bc

It is used in case-control studies as way to understanding the risk in different groups,

as it is not necessary to have a measure of the numbers initially exposed who then
develop disease. Note that when the disease is rare the odds ratio can be interpreted
as a relative risk, this is because as a gets smaller (a+c) approaches c and as b gets
smaller (b+d) approaches d. One other useful property of the odds ratio is that it is
reversible. i.e. the odds ratio for exposure is the same as the odds ratio for the event
of interest. To illustrate this, let us consider the odds of exposure for two groups, one
of which experienced the event of interest whilst the other group did not. Using the
terminology in Table 1

Odds of an exposure given the event = a/b

Odds of an exposure given no event = c/d

Thus:

Odds ratio of exposure (for event group compared to no event group)

a/b ad
c/d bc

which is the same as the odds ratio of an event. The relative risk does not have this
property. For the data in Table 2 the odds ratio of readmission to hospital for patients
having gastric bypass compared to patients having gastric banding = 0.107 / 0.068 =
1.57. Note that in this case, as the event of interest (readmission to hospital) is
relatively rare, the odds ratio is very similar to the relative risk of 1.52.

Page 110
Number needed to treat

This is a measure of the impact of a particular risk on patients often used in clinical
practice. It is the additional number of people that would need to be given a new
treatment in order to cure one extra person compared to the old treatment and is
calculated as the reciprocal of the absolute risk reduction = 1/ARR. Alternatively, for
a harmful exposure the number needed to treat is referred to as the number needed
to harm. It represents the additional number of individuals who need to be exposed
to the risk in order to have one extra person experience the event of interest,
compared to the unexposed group. For the gastric bypass study the number needed
to harm is 1/0.033=30.3. As it is usual to round this to the nearest whole number
above, the number needed to harm is 31. For the lymphoedema trial the number
needed to treat is 1/0.178 = 5.6. Thus the number needed to treat in order to have
one addition person without lymphoedema is 6.

However, when calculating the number needed to treat, it is important to know what
the absolute risks that it is based upon are. Even though the risk in the control group
can change dramatically, giving very different relative risks, the number needed to
treat can stay constant, as illustrated by the Figure below. Each bar represents the
results of a fictional study. For all of these studies the number needed to treat is 3.
The figure shows how the relative risk changes for different values of the risk in the
control group. When the risk in the control group is 0.01 (i.e. 1% risk of an event) the
relative risk is 34.3 whereas when the risk in the control group has increased to 0.5
(50% risk of an event) the relative risk has decreased to 1.67.

Figure 1: Figure illustrating what happens to the relative risk for different
control event rates, when the number needed to treat is fixed at 3.

100
Relative risk of benefit (log scale)

10

1
0.01 0.10 0.20 0.30 0.40 0.50

Proportion of events in the control group

Risk ladders

A risk ladder is a visual way of quantifying different risks, in comparison to each

other. It is often used in clinical practice when explaining individual risks to patients

Page 111
and enables patients to quantify their particular risk in relation to other risks. One of
the most well-known is the Calman Chart below(Calman 1996):

Table : Descriptions of risk in relation to the risk of an individual dying (D) in

any one year or developing an adverse response (A) in one year(Calman 1996)

Term used Risk range Example Risk

High = 1: 100 Transmission to susceptible household 1:1 to 1:2
contacts of measles and chickenpox (A)

Transmission of HIV from mother to child 1:6

(Europe) (A)

Gastrointestinal effects of antibiotics (A) 1:10 to 1:20

Moderate 1:100 to 1:1,000 Smoking 10 cigarettes a day (D) 1:200

All natural causes, age 40 (D) 1:850

Low 1:1,000 to All kinds of violence and poisoning (D) 1:3,300

1:10,000
Influenza (D) 1:5,000

Accident on road (D) 1:8,000

Very low 1:10,000 to 1: Leukaemia (D)

1:12,000
100,000
Playing football (D) 1:25,000

Accident at home (D)

1:26,000
Accident at work (D)
1:43,000
Homicide (D) 1:100,000

Minimal 1:100,000 to Accident on railway (D) 1:500,000

1:1,000,000
Vaccination associated polio (A)
1:1,000,000
Negligible <= 1:1,000,000 Hit by lightening (D) 1:10,000,000

Release of radiation by nuclear power 1:10,000,000

station

Points to consider when communicating risk

Individuals who do not deal with numbers and data regularly can often struggle to
understand measures of risk, and this case it can be useful to express risks in terms
of natural frequencies rather than percentages. Thus if we assume that the success
rate following a single cycle of IVF is about 33% then it is more easily understood by
stating that of 100 women undergoing treatment 33 will become pregnant. In
addition, how a risk is perceived will depend upon how it is presented. Relative risks
are often presented, but these cannot but properly understood without reference to
the baseline risks involved. Whilst a relative risk of 2 might sound large, if the

Page 112
underlying baseline risk is 1 in 10,000 and this increases to 2 in 10,000, then this will
represent a very different risk to an individual than if the baseline risk were 1 in 10
compared to 2 in 10. They both have the same relative risk, but the risk to an
individual is very different. When presented with a risk expressed in relative terms it
is always useful to know what the baseline risk is. A good further description of both
risk ladders and the communication of risk can be found in an article by Edwards et
al(Edwards et al. 2002).

It is also worth bearing in mind that in all the data that have been presented in this
tutorial, no other factors have been taken into account. This is particularly important
when considering the gastric bypass/gastric band data. As these are data from an
observational cohort study and not from a randomised controlled trial, it may be that
the patients who underwent gastric bypass were different to the patients who
underwent gastric banding and these other differences could explain the difference in
risk of readmission by 28 days, rather than any underlying risk associated with the
actual procedure.

Page 113
References
References

Marital status of the UK population, by sex, 2001.

www.statistics.gov.uk/STATBASE/Expodata/Spreadsheets/D7680.xls . 17-6-2005.
Ref Type: Electronic Citation

Altman, D.G. 1991. Practical Statistics for Medical Research London, Chapman & Hall.

Altman, D.G. & Bland, J.M. 1996. Presentation of numerical data. British Medical Journal,
312, 572

Altman, D.G. & Bland, J.M. 1999. Treatment allocation in controlled trials: why randomise.
British Medical Journal, 318, 1209

Altman, D.G., Machin, D., Bryant, T., & Gardner, M.J. 2000. Statistics with Confidence, 2nd
ed. London, BMJ Books.

Armitage, P., Berry, P.J., & Matthews, J.N.S. 2002. Statistical Methods in Medical Research,
4 ed. Oxford, Blackwells.

Bigwood, S. & Spore, M. 2003. Presenting Numbers, Tables and Charts Oxford, Oxford
University Press.

Bird, D. Methodology for the 2004 annual survey of hours and earnings. Labour market
trends. http://www.statistics.gov.uk/articles/nojournal/ASHEMethod_article.pdf , 457-464.
2004. Office for National Statistics. 17-6-2004.
Ref Type: Electronic Citation

Boogert, A., Manhigh, A., & Visser, G.H.A. 1987. The immediate effects of chorionic villus
sampling on fetal movements. American Journal of Obstetrics and Gynaecology, 157, 137-
139

Bradford Hill, A. 1990. Memories of the British streptomycin trial: the first randomised
clinical trial. Controlled Clinical Trials, 11, 77-79

Burns, E.M., Naseem, H., Bottle, A., Lazzarino, A.I., Aylin, P., Darzi, A., Moorthy, K., &
Faiz, O. 2010. Introduction of laparoscopic bariatric surgery in England: observational
population cohort study. British Medical Journal, 341, c4296

Calman, K. 1996. Cancer: science and society and the communication of risk. British Medical
Journal, 313, 799-802

Campbell, M.J., Machin, D., & Walters, S.J. 2007. Medical Statistics: a textbook for the
health sciences, 4 ed. Chichester, Wiley.

Cox, I.M., Campbell, M.J., & Dowson, D. 1991. Red blood cell magnesium and chronic
fatigue syndrome. Lancet, 337, 757-760

Day, S.J. & Altman, D.G. 2000. Blinding in clinical trials and other studies. British Medical
Journal, 321, 504

Page 114
Edwards, A., Elwyn, G., & Mulley, A. 2002. Explaining risks: turning numerical data into
meaningful pictures. British Medical Journal, 324, 827-830

Ehrenberg, A.S.C. 2000. A primer in data reduction Chichester, John Wiley & Sons.

Freeman, J.V. & Campbell, M.J. 2006. Basic test for continuous data: Mann-Whitney U and
Wilcoxon signed rank sum tests. Scope, 15, (4)

Freeman, J.V. & Julious, S.A. 2005a. Describing and summarising data. Scope, 14, (3)

Freeman, J.V. & Julious, S.A. 2005b. The Normal Distribution. Scope, 14, (4)

Freeman, J.V. & Julious, S.A. 2005c. The visual display of quantitative information. Scope,
14, (2) 11-15

Freeman, J.V. & Julious, S.A. 2006a. Basic tests for continuous Normally distributed data.
Scope, 15, (3)

Freeman, J.V. & Julious, S.A. 2006b. Hypothesis testing and estimation. Scope, 15, (1)

Freeman, J.V. & Julious, S.A. 2007. The analysis of categorical data. Scope, 16, (1) 18-21

Freeman, J.V. & Young, T.A. 2009. Correlation coefficient: association between two
continuous variables. Scope 31-33

Huff, D. 1991. How to lie with statistics London, Penguin Books.

Julious, S.A. & Mullee, M.A. 1994. Confounding and Simpson's paradox. British Medical
Journal, 308, 1408-1481

Julious, S.A. & Zariffa, N. 2002. ABC of pharmaceutical trial design. Phamaceutical
Statistics, 1, (1) 45-54

Kroenke, K., Spitzer, R.L., Williams, J.B., Monahan, P.O., & Löwe, B. 2007. Anxiety
disorders in primary care: prevalence, impairment, comorbidity and detection. Annals of
Internal Medicine, 146, (317) 325

Lacomba, M.T., Sanchez, M.J., Goni, A.Z., Merino, D.P., Moral, O., Tellez, E.C., &
Mogollon, E.M. 2010. Effectiveness of early physiotherapy to prevent lymphoedema after
surgery for breast cancer: randomised, single blinded, clinical trial. British Medical Journal,
340, b5396

Lang, T.A. & Secic, M. 1997. How to report statistics in medicine Philadelphia, American
College of Physicians.

Medical Research Council 1948. Streptomycin treatment of pulmonary tuberculosis. British

Medical Journal, 2, (769) 782

Morrell, C.J., Walters, S.J., Dixon, S., Collins, K., Brereton, L.M.L., Peters, J., & Brooker,
C.G.D. 1998. Cost effectiveness of community leg ulcer clinic: randomised controlled trial.
British Medical Journal, 316, 1487-1491

Nisar, A., Morris, M.W.J., Freeman, J.V., Cort, J.M., Rayner, P.R., & Shahane, S.A. 2007.
Subacromial bursa block is an effective alternative to Interscalene block for postoperative

Page 115
pain control after arthroscopic subacromial decompression - a randomised trial. Journal of
Shoulder and Elbow Surgery, In press,

Simpson, E.H. 1951. The interpretation of interaction in contingency tables. Journal of the
Royal Statistical Society, Series B (2) 238-241

Swinscow, T.D.V. & Campbell, M.J. 2002. Statistics at square one, 10 ed. London, BMJ
Books.

Tufte, E.R. 1983. The visual display of quantitative information Cheshire, Connecticut,
Graphics Press.

Williams, E.J. 1949. Experimental designs balanced for the estimation of residual effects of
treatments. Australian Journal of Scientific Research, Series A (2) 149-168

Yoshioka, A. 1998. Use of randomisation in the Medical Research Council's clinical trials of
streptomycin in pulmonary tuberculosis in the 1940s. British Medical Journal, 317, (1220)
1223

Page 116