Correlation

1 Introduction
We are often interested in the relationship between two variables.

• Do people with more years of full-time education earn higher salaries?

• Do factories with more safety officers have fewer accidents?

Questions like this only make sense if the possible values of our variables have a natural
order. The techniques that we look at in this handout assume that variables are measured
on a scale that is at least ordinal. In discussing Pearson’s correlation coefficient, we shall
need to go further and assume that we have interval scale data - i.e. that equal intervals
correspond to equal changes in the characteristic that we are trying to measure.

2 Plotting the data

The first step in looking for a correlation is to draw a scatterplot of the data. Figure 1
shows four examples1 .

2.1 Interpreting a scatterplot

These are some of the points to look for.

• How strong is the relationship?

– In Figure 1, the relationship between gas consumption and outside temperature

is very strong, while the relationship between Educational level and Crime rate
is much weaker.

• Is the relationship increasing or decreasing?

– In the ‘Gas’ example, higher outside temperatures are associated with lower
gas consumption, but in the ‘Ice cream’ example, higher mean temperatures
go with higher levels of ice cream consumption.
1

1
 r = − 0.971, ρ = − 0.958, τ = − 0.853 r = 0.776, ρ = 0.829, τ = 0.644

0.25 0.30 0.35 0.40 0.45 0.50 0.55

7

Ice cream (pints per head)

6
Gas consumption

5
4
3

0 2 4 6 8 10 −20 −15 −10 −5

Outside temperature (°C) Mean temperature (°C)

r = 0.323, ρ = 0.256, τ = 0.169 r = 0.935, ρ = 0.982, τ = 0.907

200

2.0
Direct current output
150

1.5
Crime rate

100

1.0
0.5
50

90 95 100 105 110 115 120 4 6 8 10

Educational level Wind velocity (mph)

Figure 1: Scatterplots showing strong and weak relationships

• Is the relationship roughly linear?

– The plot in the top left of Figure 1 shows a clear linear pattern, while the plot
in the bottom right suggests a non-linear relationship with the initial steep
slope leveling off as the wind speed increases.

• What is the slope of the relationship?

– Is an increase in one variable associated with a small, or a large, increase in

the other one? For example, factories with more safety officers may have fewer
accidents, but is the reduction in accidents large enough to justify the cost of
the additional safety officers?

• Are there any outliers?

– Figure 2 shows a plot of Police expenditure per head against Population size

2
for 47 US states2 . At first glance, there seems to be an increasing relationship,
with larger states spending more per head on policing. But if you cover up the
two points at the top right of the plot, the correlation seems to disappear. The
evidence for a correlation comes almost entirely from these two points, so we’ll
need to check our data source to make sure that the points really are correct.

r = 0.514, ρ = 0.301, τ = 0.219

160
140
Police expenditure per head

120
100
80
60
40

0 50 100 150

State population size

Figure 2: Effect of outliers

3
3 Correlation coefficients
A correlation coefficient gives a numerical summary of the degree of association between
two variables - e,g, to what degree do high values of one variable go with high values of
the other one?
Correlation coefficients vary from -1 to +1, with positive values indicating an increasing
relationship and negative values indicating a decreasing relationship.
We focus on two widely used measures of correlation - Pearson’s r and Kendall’s τ .

• Pearson’s coefficient

– measures degree to which a relationship conforms to a straight line

• Kendall’s coefficient

– measures degree to which a relationship is always increasing or always decreas-

ing

Spearman’s rank correlation coefficient, ρ behaves in much the same way as Kendall’s τ ,
but has a less direct interpretation.

3.1 Which coefficient should I use?

• Interval scale data and interested in linear relationships - e.g. wish to build linear
model

– Use Pearson’s coefficient

• Interested in any increasing/decreasing relationship

– Use Kendall’s coefficient

3.2 Pearson’s coefficient

Suppose we have n data pairs (xi , yi )
Pearson’s correlation coefficient is given by
Pn
i=1 (xi − x)(yi − y)
r = qP Pn
n
i=1 (xi − x)2 i=1 (yi − y)2

. . . where x and y are the means of the x and y values.

In practice, we always use statistical software to do the calculations.
Looking back at Figure 1, notice how the absolute size of the coefficient drops towards zero
as we get more and more scatter. While Pearson’s r is good at measuring the strength of

4
a linear association, it can be quite misleading in the presence of curvature. Look at the
wind turbine data in the bottom right plot of Figure 1. Pearson’s r is 0.935, suggesting a
strong linear association, but a linear model would clearly not be sensible here.
Because Pearson’s r is based on the idea of linearity, it only makes sense for data that is
measured on at least an interval scale. For ordinal data, use Kendall’s τ or Spearman’s ρ.

3.4 Modification of Kendall’s coefficient for tied data

If some of the pairs of observations are tied, Kendall’s coefficient cannot reach the limiting
values of ±1 even if all untied pairs are concordant (discordant). This is a particular
problem in the analysis of contingency tables, where there will usually be a large number
of ties. Kendall proposed the following as an alternative to the simpler coefficient defined
above.

C −D
τb = q
(n(n − 1)/2 − tx )(n(n − 1)/2 − ty )

5
. . . where tx is the number of tied x values and ty is the number of tied y values.
This version of Kendall’s τ is the one used by SPSS.

3.5 Interpreting a correlation coefficient

It’s easy to misinterpret a correlation coefficient. These are some of the points to watch.

• A correlation coefficient can be badly affected by one or two outlying observations.

For the ‘police expenditure’ data in Figure 2, the value of Pearson’s r is 0.514, but
if the two points at the top right of this plot are removed, the correlation drops to
0.237. Always look at a scatter plot before calculating a correlation coefficient!

• Correlation is not the same as causality. For example, factories with more safety
officers may have fewer accidents, but this doesn’t prove that the variation in acci-
dent levels is attributable to the provision of safety officers. The correlation may be
a spurious one induced by another factor such as the age of the factory.
One possible approach here is to use partial correlation. We ’adjust’ our two vari-
ables to remove any variation that can be accounted for by our third variable (age
of factory) and then look at the correlation between the two adjusted variables.

• Even if a relationship is genuine, a strong correlation doesn’t necessarily imply that

a change in one variable will produce a large change in the other one. The two
sets of data shown in Figure 3 give the same correlation coefficient, but say quite
different things about the effect of engine capacity on fuel economy.

Large effect (r = 0.832) Small effect (r = 0.832)

45

45
40

40
Fuel economy

Fuel economy
35

35
30

30
25

25

1500 2000 2500 3000 1500 2000 2500 3000

Engine capacity Engine capacity

Figure 3: Correlation and size of effect

• Correlation coefficients are subject to sampling variation and may give a misleading
picture of the correlation in the population we’re sampling. We can quantify the
uncertainty in an estimate of a correlation by quoting a confidence interval, or

6
 range of plausible values. For the ’Ice cream’ data in Figure 1, the 95% confidence
interval for Pearson’s r is 0.576 to 0.888, so we can be fairly sure that the population
coefficient lies in this range.
For details of how to calculate confidence intervals for correlation coefficients, see
Howells(1994) and Hollander(1999).

3.6 Testing for zero correlation

Most statistical software packages allow us to check whether a sample correlation is com-
patible with zero correlation in the population we’re sampling. The test that is carried
out here first assumes that the population correlation is zero and calculates the chance of
obtaining a sample correlation as large or larger in absolute size than our observed value
- this chance is given as the p value. If the p value is very small, we conclude that our
sample correlation is probably incompatible with zero correlation in the population.
The limitation of a test for zero correlation is that it doesn’t tell us anything about the
size of the correlation. A correlation can be nonzero, but too small to be of any practical
interest. For example, if we test for zero correlation with the data in the plot in the
bottom left of Figure 1, we obtain p value of 0.027, which gives strong evidence for a
nonzero correlation. But would a relationship as weak as this be of any practical interest?