CORRELATION

THE DEGREE OF
RELATIONSHIP BETWEEN
VARIABLES
 CORRELATION
• Correlation examines the strength of a
connection between two characteristics
belonging to the same individual, or event
or equipment
• The concept of correlation does not
include the proposition that one thing is
the cause and the other the effect
• We merely say that two things are
systematically connected
 RELATIONSHIPS
• Two variables can be positively correlated - an increase
in one variable coincides with an increase in another
variable, e.g. the more electricity used the higher the
power bill.

• A negative correlation - when one variable increases as

the other decreases, e.g. as price increases, demand
decreases.

• A zero or random correlation - when variations in two

variables occur randomly, e.g. number of accountants
graduating per year with total annual attendance at
national football league matches
 THE CORRELATION
COEFFICIENT (R)
• The Correlation is measured by the
correlation coefficient which is usually
designated as ‘r’.

• Correlations (r) range from +1.00 perfect

positive to -1.00 perfect inverse with a
midpoint 0.00 indicating absolute
randomness
GUIDE TO INTERPRETING CORRELATION SIZE

0.90 - 1.00 -Very high correlation - Very strong

relationship

0.70 - 0.90 -High correlation - Substantial relationship

0.40 - 0.70 -Moderate correlation - Moderate relationship

0.20 - 0.40 -Low correlation - Weak relationship

0.00 - 0.20 -Slight correlation - Relationship small and

random
 Perfect Positive Correlation
12

10

0
Y

0 2 4 6 8 10 12

X
Fig 15.1(a) perf ect positive correlation +1.00
 Perfect Negative Correlation
12

10

2
Y2

0
0 2 4 6 8 10 12

X
Fig 15.1(b) Perf ect inverse (negative) correlation -1.000
 High Positive Correlation
12

10

2
Y3

0
0 2 4 6 8 10 12

X
Fig 15.1(c) positive correlation +0.867
 Moderate Negative Correlation
12

10

2
Y4

0
0 2 4 6 8 10 12

X
Fig 15.1(d) negative correlation -0.576
 Random Correlation
12

10

2
Y5

0
0 2 4 6 8 10 12

X
Fig 15.1(e) random correlation -0.079
 DIRECTION AND STRENGTH OF
CORRELATION COEFFICIENTS

• The direction of relationship is indicated by

the mathematical sign + or –

• The strength of relationship is represented

by the absolute size of the coefficient -
how close it is to +1.00 or -1.00.
 CORRELATION TESTS
There are many correlation tests designed
for particular conditions such as whether
data is ordinal or interval or combinations
of the two. The major ones are:
• Parametric Tests of Correlation
– Pearson Product Moment Correlation or ‘r’
– Used with scale data
• Non-Parametric Test of Correlation:
– Rank Order Correlation (Spearman's Rho)
– Cramer’s V, Kendall’s tau and phi, biserial
correlation, tetrachoric correlation
 ASSUMPTIONS OF
CORRELATIONAL ANALYSIS
1. Data must come from related pairs: if you have
a score on X there must also be a score on Y
from the same person/equipment/event
2. The measurement must be interval or ratio
(scale data on SPSS) or ordinal
3. Each set of scores should be approximately
normal
4. The relationship between the two sets of
scores should be linear or approximately so
5. The variability of scores for one variable should
be roughly similar at all values of the other
variable (homoscedasticity)
 CALCULATING PEARSON'S ‘r’
We shall test whether there are significant relationships
between
1. starting salary and current salary,
2. current salary and age, and
3. age and number of days absent.

The Ho for each of these three correlation tests is that

there is no statistically significant relation between each
pair of variables.

SPSS can carry out a number of correlation tests

simultaneously although as the number increases the
output table of inter-correlations becomes quite large
and unwieldy on the computer screen as it provides
inter-correlations between every variable entered.
 PEARSON CORRELATION SPSS
• Chap 13 data set A contains a number of variables
concerning the work place.
1. Click Analyze, select Correlate and choose Bivariate.

2. Select your variables (e.g. current salary, start salary, age and
number of days absent in 2006) and click the arrow button which
places them in the Variables box.

3. The Pearson option is the default and ensure Flag Significant

Correlations is checked.

4. Select OK to close the Bivariate Correlations dialogue box and

produce the output.

5. Should you wish, you can also obtain Means and Standard
Deviations by clicking Options then selecting these items in the
Statistics box.
Pearson Correlation SPSS
PEARSON CORRELATION SPSS
• Additionally, you should create a
scattergraph to display the relationship
visually.
• It lets you see whether you have a linear
or nonlinear relationship. Correlation
assumes a linear or closely linear
relationship.
• If the scattergraph shows a curvilinear
shape then you cannot use Pearson.
PRODUCING A SCATTERGRAPH
• 1. Click Graph and then Scatter.
• 2. Simple is the default mode and will provide a
single scattergraph for one pair of variables.
Click Define. (If you wish to obtain a number of
scattergraphs simultaneously click on Matrix
then Define.)
• 3. Move two variables (e.g. current and starting
salary) into the box with the arrow button, then
select OK to produce the scattergraph
• 4. With a correlation it does not really matter
which variable you place on which axis
 Scatterplot using SPSS
Each point on the scatterplot represents the intersection for a pair
of data. The Y axis represents values of the dependent variable
(annual current salary) and the X axis the independent (annual
starting salary). Approximate linearity is demonstrated
100

90

80
annual current salaryin$1000's

70

60

50

40

30
10 20 30 40 50

annual starting salary in $1000's

 Example of a Curvi-linear
Relationship
High
Positive
correlation
area

Level Negative
of correlation
Perfor area
mance

Low

Stress Levels
Low High
 Output Inter-correlation Table
Pearson Correlations
annual
no.of days starting annual
absent in salary in current salary
2006 $1000's in $1000's age
no.of days absent in Pearson Correlation
1 .070 .027 -.148
2006
Sig. (2-tailed) . .666 .866 .364
N 40 40 40 40
annual starting salary Pearson Correlation
.070 1 .735(**) .056
in $1000's
Sig. (2-tailed) .666 . .000 .733
N 40 40 40 40
annual current salary Pearson Correlation
.027 .735(**) 1 .354(*)
in $1000's
Sig. (2-tailed) .866 .000 . .025
N
40 40 40 40

age Pearson Correlation -.148 .056 .354(*) 1

Sig. (2-tailed) .364 .733 .025 .
N 40 40 40 40
** Correlation is significant at the 0.01 level (2-tailed).
* Correlation is significant at the 0.05 level (2-tailed).
 INTERPRETING PEARSON OUTPUT SPSS

• Only a correlation matrix is produced. The

diagonal of this matrix (from top left to bottom
right) consists of each variable correlated with
itself, obviously giving a perfect correlation of
1. No significance level is quoted for this
value.

• Each result is tabulated twice at the

intersections of both variables. The exact
significance level is given to three decimal
places.

• Forty pairs of scores were used to obtain the

correlation coefficient.
 INTERPRETING PEARSON OUTPUT
SPSS

• The correlation between starting and current

salary is +0.735; this is significant at the .01
level, i.e high starting salaries are related to high
current salaries and low with low
• The correlation between current salary and age
is +0.354; this is significant at the .05 level, i.e a
weak relationship between increasing salary and
increasing age.
• The correlations between age and absence, and
age and annual starting salary are not
significant and both are close to a random
relationship, i.e. they are not related at all
 SPEARMAN’S RANK ORDER CORRELATION

• This correlation is typically used when there are only a

few cases (or subjects) involved and the data are
already ordinal or can be changed into ranks.
• This correlation is usually designated as 'rho' to
distinguish it from Pearson's ‘r’. It is based on the
differences in ranks between each of the pairs of ranks.
E.g. if a subject is ranked first on one measure but only
fifth on another the difference (d) = 1 - 5 = 4.
• Rank order correlation follows the same principles as
outlined earlier and ranges from +1 to -1.
• While rho is suitable only for ranked (or ordinal) data,
interval data can also be used provided it is first ranked.
• However, rho will always provide a lower estimate of
correlation than r because the data is degraded
 EXAMPLE OF SPEARMAN RHO
• To demonstrate this statistic, a sample of 40 employees provided
data on the relationship between their sales figures and current
salary. Access SPSS Chap13 Data File A
• How to proceed
• If scores are not already in ranks you need to transform them into
ordinal data. To do this:
• 1. Select Transform and from the drop-down menu, click on
Rank Cases
• 2. Select the arrow button to move the variable (sales 3) into the
Variables : text box
• 3. Select the Largest value button in the Assign Rank 1 to area
(Fig 15.8)
• 4. Click OK and a new variable is created on the data view screen
reflecting the ranking of the original variable. This new variable will
carry the same name as the original but be preceded by a letter ‘r’ to
designate it as a ranked variable (e.g. ‘sales3 will become rsales3’,
‘age’ will become ‘rage’, etc;) Rank all interval or scale variables to
be used with Spearman this way.
SPSS EXAMPLE OF RANKING
VARIABLES
 Spearman’s Rho using SPSS
• 1. Click Analyze, select Correlate and
choose Bivariate.
• 2. Select your variables (in this case rank
of sales3 and rank of SalNow) and move
them to the Variables box.
• 3. Select Spearman in the Bivariate
Correlations dialogue box and de-select
Pearson the default statistic.
• Select OK to close the Bivariate
Correlations dialogue box and display the
output (shown in the following slide).
Spearman’s Rho SPSS
 Spearman’s Rho Printout SPSS

RANK of RANK of
SALNOW SALES3
Spearman's rho RANK of SALNOW Correlation
1.000 .010
Coefficient
Sig. (2-tailed) . .953
N 40 40
RANK of SALES3 Correlation
.010 1.000
Coefficient
Sig. (2-tailed) .953 .
N 40 40
 INTERPRETATION OF SPEARMAN’S RHO
PRINTOUT

• Spearman's rho is printed in a matrix like

that for Pearson with 1 in the top right
diagonal and the results replicated in the
top right to bottom left diagonal.
• Spearman's correlation of +0.010
between the ranks for current salary and
sales performance at period three is
close to random and not statistically
significant ( p = .953)
• The number of cases on which that
correlation was based is 40.
 COEFFICIENT OF DETERMINATION

• The Coefficient of Determination (r2) represents the proportion

of variation in one variable which is explained by the other.

• r2 is always positive and usually reported as a percentage.

• It is calculated by squaring the r value.

e.g. r2 = .9852
= .970

Therefore, 97% of the variance in variable Y is explained by the

variance in variable X.
Coefficient of Determination

– If the r between variable X and variable Y = 0,

then the coefficient of determination = 0 or
0%. None of the factors accounting for
variability is common to both variables

Total Total
Variance Variance
X Y
 COEFFICIENT OF DETERMINATION

• If the (r) between variable X and variable Y = 0.8, then

the coefficient of determination = 0.82(100) = 0.64(100)
or 64%. Therefore 64% of the factors accounting for
variability are common to both variables, or in other
words, 64% of the variance in one variable is predictable
from the variance in the other.

X Y Shared
X Variance Y
64%
 COEFFICIENT OF DETERMINATION
• If the correlation (r) between variable X and
variable Y = 1, then the coefficient of
determination = 12 = 1 x 100 = 100%. 100%
of the factors accounting for variability are
common to both variables.

XY
Shared
Variance
100%
 PROBLEMS IN INTERPRETING A
CORRELATION
• Whenever a correlation is computed from
values that do not represent the full range
of the distribution caution must be used in
interpreting it.
• A high positive correlation can be
obscured if only a limited range of data is
obtained (following slide)
• An overall zero ‘r’ can be inflated to a
positive one by restricting the range too
(following slide)
Restricted Range producing Zero ‘r’
Restricting Range to Produce +ve
‘r’
 PARTIAL CORRELATION
• Partialling out a variable is used when you wish
to eliminate the influence of a third variable on
the correlation between two other variables. It
simply means controlling the influence of that
variable.
• Other terms for partialling out are ‘holding
constant’, and ‘controlling for’.
• The partial correlation coefficient, which, like
other correlations takes values between -1 and
+1, is essentially an ordinary bivariate
correlation, except that some third variable is
being controlled for.
 EXAMPLE OF PARTIAL CORRELATION
• We will illustrate this using Chap 13 data set with
data from 40 employees on their age, their
current salary, and their starting salary.

• A previous study had shown a significant

Pearson’s ‘r’ of +0.735 between starting salary
and current salary but this could be an artifact
of their age since this likely impacts on both.

• A partial correlation was computed to ‘hold

constant’ or ‘partial out’ the variable of age in
the relationship between starting and current
salary to check whether this assumption is
correct.
 PARTIAL CORRELATION SPSS
• 1. Select Analyze and click on Correlate.
• 2. Choose Partial which brings the Partial
Correlations box into view.
• 3. Select the variables that represent starting salary
and current salary and move them into the variables box.
• 4. Select the age variable and move it into the
Controlling for box
• 5. Click the two- tail option in the Test of
Significance box.
• 6. To obtain the correlation between starting salary and
current salary without holding age constant click
Options and select Zero order correlations
• 7. Click Continue then OK. Two tables are produced
SPSS EXAMPLE OF PARTIAL CORRELATION
 INTERCORRELATIONS BETWEEN
VARIABLES
P A R T I A L C O R R E L A T I O N C O E F F I C I E N T S - - -

Zero Order Partials

STARTSAL SALNOW AGE

STARTSAL 1.0000 .7353 .0556

These inter-variable
( 0) ( 38) ( 38) correlations are the
P= . P= .000 P= .733 normal Pearson ‘r’s
before partialling out
SALNOW .7353 1.0000 .3535 age.
( 38) ( 0) ( 38)
P= .000 P= . P= .025

AGE .0556 .3535 1.0000

( 38) ( 38) ( 0)
P= .733 P= .025 P= .
 Intercorrelations with Age partialled
Out
PARTIAL CORRELATION COEFFICIENTS ---

Controlling for.. AGE

STARTSAL SALNOW
These correlations now
reflect the partialling out
STARTSAL 1.0000 .7662 of age. Note the
( 0) ( 37) correlation has increased
marginally from +0.735 to
P= . P= .000 +0.766

SALNOW .7662 1.0000

( 37) ( 0)
P= .000 P= .
 INTERPRETATION OF PARTIAL
CORRELATION PRINTOUT
• The first table provides usual bivariate correlations
between starting salary and current salary now
without age held constant. This displays a correlation
of +0.7353, p < 0.001.
• The second table provides the partial correlation
coefficient, with age held constant and the p value on a
diagonal format as usual.
• The partial correlation between starting salary and
current salary with age held constant is +0.7662, with
p < 0.001 which indicates a stronger significant
relationship than the zero order one of +0.7353. Age
has a small but important effect in moderating the
correlation between starting salary and current salary.
• The null hypothesis that there is no significant
relationship between starting salary and current salary
with age partialled out must be rejected.