What is the correlation coefficient?

Linear means straight line. Correlation means co-relation, or the degree that two variables "go
together". Linear correlation means to go together in a straight line. The correlation coefficient
is a number that summarizes the direction and degree (closeness) of linear relations between two
variables. The correlation coefficient is also known as the Pearson Product-Moment Correlation
Coefficient. The sample value is called r, and the population value is called  (rho). The
correlation coefficient can take values between -1 through 0 to +1. The sign (+ or -) of the
correlation affects its interpretation. When the correlation is positive (r > 0), as the value of one
variable increases, so does the other. For example, on average, as height in people increases, so
does weight.

1 60 102

2 62 120

3 63 130

4 65 150

5 65 120

6 68 145

7 69 175

8 70 170

9 72 185

10 74 210

Example of a Positive Correlation

If the correlation is positive, when one variable increases, so does the other.

If a correlation is negative, when one variable increases, the other variable descreases. This
means there is an inverse or negative relationship between the two variables. For example, as
study time increases, the number of errors on an exam decreases.
N Study # Errors
Time on test
Min.
25
1 90
28
2 100
20
3 130
20
4 150
15
5 180
12
6 200
13
7 220
10
8 300
8
9 350
6
10 400

Example of a Negative Correlation

If the correlation is negative, when one variable increases, the other decreases.

If there is no relationship between the two variables, then as one variable increases, the other variable
neither increases nor decreases. In this case, the correlation is zero. For example, if we measure the SAT-
V scores of college freshmen and also measure the circumference of their right big toes, there will be a
zero correlation.
Note that as either toe size or SAT increases, the other variable stays the same on average.

The conceptual (definitional) formula of the correlation coefficient is:

(1.1)

where x and y are deviation scores, that is,

and

SX and SY are sample standard deviations, that is,

This says that the correlation is the average of cross-products (also called a covariance)
standardized by dividing through by both standard deviations.

Another way of defining the correlation is:

 (1.2)

where zx is X in z-score form,

zy is Y in z-score form,

And  and N have their customary meaning. This says that r is the average cross-product of z-
scores. Memorize these formulas (1.1 and 1.2). These formulas are equivalent. Recall that

In equation 1.1, x is the deviation score, and SX appears in the denominator. In equation 1.2, SX is
divided through and the deviation score appears as zX.

Sometimes you will see these formulas written as:

and

These formulas are correct when the standard deviations used in the calculations are the
estimated population standard deviations rather than the sample standard deviations, that is,
when

so the main point is to be consistent. Either use N throughout or use N-1 throughout.

Example: Height and Weight Revisited

N Ht Wt Zht Zwt Zh*Zw

 1 60 102 -1.4974 -1.4344 2.1479

2 62 120 -1.057 -0.90424 0.95578

3 63 130 -0.83679 -0.6097 51019

4 65 150 -0.39637 -0.02062 0.00817

5 65 120 -0.39637 -0.90424 0.35842

6 68 145 0.26425 -0.16789 -0.04435

7 69 175 0.48446 0.715741 0.34674

8 70 170 0.70466 0.56846 0.40058

9 72 185 1.14508 1.01028 1.15685

10 74 210 1.58548 1.74663 2.76927

66.8 150.7 0 0 .96*

S(N-1) 4.54117 35.9111 1 1

* The value of .96 comes from dividing the sum of products by 9 rather than 10. I used (N-1)
throughout.

Points to notice: The mean height is 66.8 inches, the SD is 4.54 inches. The first height is 60
inches, which is about 1.5 sample standard deviations below the mean, or a z-score of -1.4974.
The first weight is 102 pounds, which is 1.43 standard deviations below the mean weight z =
(102-150.7)/35.91 = -1.43. The product of the two z-scores is 2.1479 (-1.4974*-1.4344=2.1479).
If we average the products (actually sum and divide by N-1), we get .96, which is the correlation
coefficient.

Why does the correlation coefficient have a maximum of 1, and a min of -1? Why is the
correlation positive when both increase together?

Let's look at graphs of height and weight. First in raw scores, then in z-scores:
Raw Scores z-scores

Points to notice:

1. The pattern or plot of scores is the same in the new and original graphs, but the new mean
of height and weight are now zero, and of course the new units are standard deviations.
2. Each score is represented by a point, and the points fall into four quadrants defined by the
two means. Points in the upper right quadrant are above the mean on both X and Y. Points
in the lower right are above the mean on X and below the mean on Y. Points in the upper
left are below the mean on X and above the mean on Y. Points in the lower left are below
the mean on both X and Y.
3. The products of the z-scores will be positive if both zx and zy are positive AND if both zx
and zy are negative. Positive cross-products will correspond to points found in the upper
right and lower left quadrants. Products will be negative if one or the other, but not both,
zx and zy are negative. Negative cross products correspond to points in the upper left and
lower right quadrants. Notice that in the z-score graph I have placed a plus sign (+) in the
quadrants corresponding to the positive cross-products and a minus sign (-) in the
quadrants corresponding to the negative cross-products.
4. The correlation coefficient is the average cross- product of z-scores. If most of the points
are in the positive quadrants, the correlation will be positive. If most of the points are in
the negative quadrants, the correlation coefficient will be negative. If the points are
equally distributed in all four of the quadrants, the correlation coefficient will be zero.
5. The maximum cross-products occur when x=y, that is, when the points fall on the 45
degree straight line passing through 2 quadrants (either both positive or both negative).
This happens because then zx*zy = zx2. For example, if zx is 1, the maximum product
obtains when zy is 1 because 1*1 = 1. Any other value less than 1 results in a smaller
product, e.g., 1 * .9 = .9. Of course, if zy is greater than 1, then the product is larger than
1, e.g., 1*2 = 2, but this is smaller than if zx had been 2, the same value as zy (2*2 is 4,
which is greater than 1*2 = 2). The maximum value of the correlation coefficient (+ or -
1) occurs when the values of x and y fall on a straight line, that is when the co-relation is
perfect.
The sampling distribution of r

We don't usually know , the population correlation. We use the statistic r to estimate  and to
carry out tests of hypotheses. The most common test is whether  =0, that is whether the
correlation is significantly different from zero. For example, we might compute a correlation
between a mechanical aptitude test score and a measure of success in a mechanical training
program and test to see whether the correlation is different from zero. We can also use r to test
whether  some value other than zero, such as  =.30. For example, based on past research, we
might hypothesize that the correlation between SAT-V and GPA is .60. We can also test whether
the correlations are equal across groups (e.g., whether the correlation between SAT-V and GPA is
equal for males and females). The tests of hypotheses rest on statements of probability. For
example, we might say that the observed r of .70 is very unlikely if  =0. The statements of
probability come from sampling distributions. A sampling distribution is what you get if you take
repeated samples from a population and compute a statistic each time you take a sample. You
might recall that the (long run) average of the sampling distribution of means is the population

mean (that is , the expected value of the average of the distribution of sample means
is the parameter). Another way of saying this is that the sample mean is an unbiased estimator of
the population mean. Unbiased estimators have the property that the expected value (mean) of
the sampling distribution is the parameter. If the expected value does not equal the parameter, the
estimator is biased.

As the sample size increases, the sampling distribution of means becomes Normal in shape (this
property is known as the Central Limit Theorem). This is really cool because we can use the
Normal distribution to calculate probabilities and generate tests of hypotheses.

Unfortunately, most of the time the mean of the sampling distribution of r does not equal , nor
is the sampling distribution of r ever Normal. However, statisticians have developed ways of
dealing with this wayward statistic. But first, let's look at the sampling distribution of r. Recall
that r is bounded by +1 and -1, that is, it can take no larger or smaller values. When  =0, r is
symmetrically distributed about 0, and the mean of the sampling distribution does equal the
parameter (yay!). As  increases from zero (becomes more positive), the sampling distribution
becomes negatively skewed. As  becomes negative, the sampling distribution becomes
positively skewed.
As  approaches +1 or -1, the sampling variance decreases, so that when  is either at +1 or -1,
all sample values equal the parameter and the sampling variance is zero. The sampling variance
of r is approximately

Note that as | | approaches 1, the sampling variance approaches zero. The shape of the sampling
distribution depends on N (trust me on this one). The shape becomes increasingly Normal with
large values of N, and becomes increasingly skewed with increasing | |. As you can see, the
sampling variance, and thus the significance test, depend upon (1) the size of the population
correlation and (2) the sample size.

The Fisher r to z transformation

Fisher developed a transformation of r that tends to become Normal quickly as N increases; it's
called the r to z transformation. We use it to conduct tests of the correlation coefficient. Basically what
it does is to spread out the short tail of the distribution to make it approximately Normal, like this:
r z

.10 .10

.20 .20

.30 .31

.40 .42

.50 .55

.60 .69

.70 .87

.80 1.10

.90 1.47

Note that r to z doesn't do much until r gets large (the distribution r becomes increasingly
skewed as it increases, so z has to compensate more). The transformation is:

(1.3)

were loge is the natural logarithm, and r is the sample correlation. The transformation is
illustrated above in the table and associated graph.

Hypothesis Tests

Testing whether  =0. SAS uses the formula

and compares the result to a t distribution with (N-2) df

to return a probability value (2-tailed). Look at the p value in SAS to tell if the correlation is
significantly different from zero, that is, to test

H0:  =0.
Testing for a value of  other than zero: the one-sample z. Suppose we have done some review
work on the correlation between mechanical aptitude test scores and job performance of auto
mechanics for Chevy dealers. We know that on average, the correlation is .30. We think our new
test is better than this, so we give the test to 200 mechanics and also collect job performance data
on them. The resulting correlation is .54. We know that  =0 is not tenable for these data, but we
are going to test whether  =.30, that is, H0:  =.30. The formula to do so is:

(1.4)

which is distributed approximately as N(0,1) when the null hypothesis is true (that is, we can use
the standard normal distribution for probabilities). For our data, we have

= 4.13

Because 4.13 is greater than 1.96 (1.96 is z or standard normal for  = .05, two-tailed), we can
reject the hypothesis that  =.30 and conclude H1:   .30. We could have chosen a 1-tailed
hypothesis for our alternate H1:  >.30. In that case, we would have used z = 1.645 as our critical
value.

Testing the equality of correlations from 2 independent samples. Suppose we want to test
whether correlations computed on two different samples (NOT 2 correlations from the same
sample) are equal. For example, suppose we compute correlations between SAT-V and freshman
GPA separately for samples of male and female students at USF. We find that for males (N=150)
r = .63 and for females (N=175) r = .70. We want to know whether these are significantly
different, that is, H0:  1 =  2.

The formula is:

 (1.5)

which is distributed approximately as N(0,1) when the null hypothesis is true. For our example

= -1.18, which is not significant because it is contained in the interval

-1.96<z< 1.96. We conclude that the correlation between SAT-V and GPA is not different for
males and females at USF. Note: this is not a very powerful test (lots of times when there is a real
difference, you won't detect it with this test). Also note, we would probably be testing for
differences in regression slopes if this were for real. (How to do so will be covered later.) And,
yeah, it's phrased as accepting the null hypothesis, which would cause some people to quibble
with the diction. Never mind. There will be times, however, when you want to test for
differences in independent correlations rather than regressions, and this is how to do it. It is also
possible to test whether multiple independent rs are equal. For example, we could test whether
the correlation between SAT and GPA is equal across all the state universities in Florida. How to
do so is described in Hays and other statistics texts.

More than two independent correlations. Suppose we want to test the hypothesis that three or
more correlations from independent samples share the same population value. That is, suppose
we have k independent studies and we want to test the hypothesis that k First we
have to estimate  the common population value. To do this, we will compute the average
across the studies. But first we will use Fisther's r to z transformation. If the studies differ in
their sample sizes, then we will also need to compute a weighed average so that the studies with
the larger samples get more weight in the average than do the smaller studies. Therefore, we
compute:

,
which is just a weighted average of Fisher transformed correlations. We use (n-3) instead of n
because of the z transformation. Now we need to know how far the individual studies are from
the average. To get a handle on this, we compute:

Q is just the sum of squared deviations from the mean where the squared deviations are weighted
by their sample sizes (less 3). It turns out that when the null hypothesis that all k studies share a
single underlying value of  is true Q is distributed as chi-square with (k-1) degrees of freedom.
So we can compute Q and compare the result to a tabled value of chi-square to test for
significance. If Q is large compared to the tabled value of chi-square, then we can reject the null
hypothesis that the study correlations were drawn from a common population with a single value
of 

For example, suppose we are doing test validation work. We have computed the correlation
between bank tellers' performance on a video teller test with customer service ratings furnished
by actors in our employ who portray customers with banking problems to the tellers and then
evaluate the unsuspecting tellers on their customer service skills. Suppose we have completed
studies at three different banks, and we want to test whether the correlation between the test and
customer service ratings is the same or different across banks. Suppose we found the following
results:

(n-3)
(z-
Study r n z (n-3)z zbar (z-
zbar)2
zbar)2
1 .2 200 .2 39.94 .41 .0441 8.69
2 .5 150 .55 80.75 .41 .0196 2.88
3 .6 75 .69 49.91 .41 .0784 5.64
sum 425 170.6 17.21

The three studies' outcomes or sample rs were .20, .50 and .60. The study sample sizes are
shown under the column labeled n. We transform r to z in the next column, which is labeled z.
To find the weighted average, we first multiply each z by (n-3), the study sample size less three.
When we add the weighted zs, we find the sum is 170.6 The total sample size is 425, which we
reduce by nine (reduce n by three for each of the three studies) for a value of 416. Our weighted
average, zbar, is 170.6/416 or .41. To find the value of Q, we subtract zbar from each z and
square (z-zbar)2. Finally, we weight each squared term by its sample size less three. The
weighted, squared entries are shown in the final column of the table. The sum of the entries in
the rightmost column of the table is our value of Q, 17.21. We compare this value to chi-square
with (k-1) = 2 (in our case) degrees of freedom. The value of chi-square at p = .05 is 5.99, so we
can reject the null hypothesis that all studies were drawn from a common population. It appears
that the correlation between test scores and customer services is larger at some banks than at
others.
Testing for equality of dependent correlations. Suppose you want to test whether two correlations
taken from the same sample are equal, that is, you want to test the equality of two elements from
the same correlation matrix. There is a literature on this, and there are several available tests.
There are two types of test, namely  12 =  13 or  12 =  34. Note that in the first case, there is one
variable in common, and that in the second case, there is no variable in common. An example of
the first type would be a test of whether the correlation between cognitive ability scores and
performance in a class on U.S. history is equal to the correlation between cognitive ability scores
and performance in a class on psychology. An example of the latter would be whether the
correlation between scores on a cognitive ability test and performance on a class test in
psychology is equal to the correlation between personality test scores and scores on an interview
used to determine a person's suitability for personal counseling as a therapist (whew, what a long
sentence). Both types of test are described by Steiger (1980, Psychological Bulletin). For the
following hypothesis,

Bobko (1995) recommends the "Hotelling-Williams test." Steiger also allows that is test is
reasonable for the above hypothesis.

The Hotelling-William test is:

where

and

For example, suppose we have administered a final examination in a statistics class (1). In
addition, we have two other measures. One of the measures is the SAT Quantitative score (2),
and the other is self-efficacy questionnaire (3) that asks each student how proficient they think
they are at various kinds of stats problems. We want to know whether the final exam score is
more highly correlated with the SAT-Q or the self-efficacy scale. That is, we want to know
whether

Suppose that we have 101 students in our class. Further suppose that r12 (the correlation between
the exam and the SAT) is .4, r13 (exam and self-efficacy) is .6 and r23 is .3.
then

and

Therefore

If we find the critical value for t with 98 degrees of freedom we find the value 1.98. We can
therefore reject the null that the correlations are equal in the population. Self efficacy is more
highly correlated with our exam score than was the SAT-Q.

It is also possible to test the hypothesis that  12 =  34. I just finished a study in which we
administered a self-efficacy questionnaire (how good are you at problems involving the mean,
median or mode?) and an exam (actual problems on the mean, median, or mode) to students in
research methods twice, once at the beginning of the semester, and once at the end of the
semester. We would expect both self-efficacy and exam performance to increase with
instruction. We also expect that students will become increasingly well calibrated over the
course of instruction. That is, we expect them to get a better handle on how good they actually
are during the semester as they receive feedback on their work. Suppose we have 100 students,
and our results look like this:

1 2 3 4

1 Self Efficacy Pretest 1

2 Exam Pretest .40 1

3 Self Efficacy Posttest .30 .45 1

4 Exam Posttest .10 .35 .70 1

If students become better calibrated, then we would expect that the correlation between exam
scores and self efficacy scores would be larger at the end of the semester than at the beginning of
the semester. We want to test whether the correlations .40 and .70 are significantly different; this
will indicate whether calibration improved significantly.

The test that Steiger (1980) recommends for this purpose can be written:
Here z12 and z34 refer to the Fisher transformed values of r, the two correlations. The value in the
denominator requires some explanation.

If we define

then

and

In our example, rbar is (.40+.70)/2 = .55. We will substitute the value of .55 for both r12 and r34
in the formula above. Carrying out a bit of arithmetic, we find that

and

The transformations of our correlations to z give us .423649 and .867301. Our test, then, is

A two-tailed test at alpha = .05 gives us a critical value of 1.96. Therefore, we can reject the null
and conclude that the correlations are different. The students became better calibrated over the
course of instruction.

Artifactual Influences on r

1. Range Restriction and Enhancement. Range restriction happens when X, Y, or both are
restricted in the values that they can take. For example, suppose we want to know the
correlation between a test such as the SAT and freshman GPA. We collect SAT test scores
from applicants and compute GPA at the end of the freshman year. If we use the SAT in
admissions and reject applicants with low scores, we will have direct range restriction
because there will be nobody in the sample with low test scores. There will also be some
restriction of range in GPA, because some of the lowest GPA folks will drop out of school
during the academic year and so be lost to our sample. Now suppose we do the same
study again (that is, correlate the SAT with FGPA) but this time a cognitive ability test
other than the SAT will be used for admissions and the SAT will not be used. When we
compute our correlation, there will be indirect restriction of range, because the SAT and
cog ability test will be correlated, so we will lose a disproportionate number of low
ability people from the sample. Some people like to point out that cognitive ability is not
important for scholastic success. They say so because in some elite schools like Harvard,
there is a minimal correlation between SAT scores and GPA. The result, however is due
to restriction of range in SAT -- only higher scoring students get in.

The graph above shows unrestricted data in which both variables are continuous and
approximately normally distributed.
 The above graph shows what happens if we delete all cases below the mean on X.

Note how the resulting r is smaller than the original .70.

2. Enhancement. A. Choose extremes. B. Mix clinical & normal samples but generalize to
one or the other.
 Here we have deleted the observations between 1 and 3. Note how the resulting r is larger
than the original .70.

3. Reliability. In psychometrics there is an old saying that "reliability sets the ceiling for
validity." Reliability describes the amount of error there is in a measure, and validity is
often expressed as a correlation coefficient. Reliability coefficients are often expressed as
the proportion of true score (the opposite of error) variance in the measure. If all the
variance is due to true scores, then the reliability is 1. If all the variance is due to error,
the reliability is 0; if half and half, the reliability is .50. (Reliability coefficients are often
estimated by alternate forms, test-retest, etc.) A formula that shows how reliability sets
the ceiling on validity is:

This says that the correlation between observed (measured) variables is equal to the correlation
between their true scores, that is, the scores that would be observed with no error times the
square root of the product of the two reliabilities, one reliability for each variable. For example,
suppose we have a correlation between true scores of .70. The reliability of X is .80 and the
reliability of Y is .80. Then the observed correlation will be
r = .70*sqrt(.8*.8) = .7*.8 = .56.

It sometimes happens that we have an observed correlation, estimates of the reliabilities of x and
y, and we want to show what the correlation would be if we had measures without error. This is
sometimes called a correlation that is disattenuated for reliability. The formula to do this is a
simple transformation of the formula above:

For example, if we had an observed correlation of .56, and a reliability of X of .8 and a reliability
of Y or .8, then we would have

r = .56/sqrt(.8*.8) = .56/.8 = .70.

Distribution Shape

The formula for r is based on the assumption of a bivariate normal distribution. This means that
both distributions are normal, and that if you choose any interval of X, the conditional values of
Y will still be normal, and if you choose any values of Y, the conditional values of X will be
normal. As long as both variables are fairly continuous and more or less Normal, r will provide a
good summary of the linear relations in the data. However, when one or both variables have a
bad shape or have only 2 or 3 values, the range of r becomes restricted. For example, if one
variable is binary and the other is continuous, r can virtually never reach a value of 1.0. Its
maximum value occurs when about half of the binary values are zero and half are 1, which is
also where the maximum variance for the binary measure occurs. The maximum value of r
decreases as the proportion of 1s in the binary variable departs from .5. For example, let's look at
a few numbers.

Person X Y1 Y2 Y3

1 1 1 0 0

2 2 2 0 0

3 3 3 0 0

4 4 4 0 0

5 5 5 0 0

6 6 6 1 0
7 7 7 1 0

8 8 8 1 0

9 9 9 1 1

10 10 10 1 1

Note that X and Y1 are identical, so the correlation between the two will be 1.0 . Note that Y2
and Y3 are simple transformations of Y1 that make it binary. Y2 transforms Y1 so that values of
5 or less are zero, while values greater than 5 are 1. Y3 transforms Y1 so that values 8 or less are
zero. The results:

X Y1 Y2 Y3

X 1

Y1 1.0 1

Y2 .87 .87 1

Y3 .70 .70 .50 1

Note that as the binary variable moves from 50 percent, the correlation decreases. Note also that
when two things correlate 1.0, they also correlate the same with everything else. Also note that
when you correlate two binary variable (Y2 and Y3), the correlation is even lower. There are
variants of the correlation coefficient that try to correct for this problem (the biserial and
tetrachoric), but there are problems with these.

Interpreting r

1. A measure of association

a. As |r| increases, the strength of linear association increases.

b. r2 (r-square[d]) is a measure of shared variance, or variance accounted for. If we find that
r2xy = .50, we can say that x "accounts for" 50 percent of the variance in y. Or we can say
that y accounts for 50 percent of the variance in x, or we can say that they share 50
percent of variance in common.
c. Although r2 is the most natural interpretation of association, r does have some
interpretations on its own. For example, in personnel selection, we can interpret the
usefulness of a test using r instead of r2 (the Taylor-Russell Tables are used for this).

2. As a measure of influence

We often talk about experimental (ANOVA) and correlational results as if we can infer cause
with one and not with the other. Actually, the statistical analysis has little or nothing to do with
inferring causality. The major thing of importance in inferring causality is the way in which data
are collected, that is, what matters is the design and not the analysis. We can assign people to
study time using random numbers with a given mean and standard deviation. For example,
before a test we could assign people to receive study time with a mean of 3 hours and a standard
deviation of 1 hour, where each person would have a different study time. We could analyze the
results with a correlation, and then talk about the influence of study time on test score. On the
other hand, we can use ANOVA to test for differences between males and females on test scores,
but we cannot say whether any difference in scores is due to sex, gender, specific experience, the
absence of chromosomes, etc. Even when we assign people to treatments, we cannot infer
causality in the way we would like to. Assignment tends to eliminate alternative explanations of
results. There are almost always alternative explanations left even after the assignment. For
example, demand characteristics of an experiment may explain the results better than the
manipulated variable (that is, the study's participants acted the way they did not because of the
manipulated variable, but because they knew they were in a study and it mattered). Remember
the Hawthorne studies.

It is useful to think about observed, measured variables as being composites made up of several
things, only one of which we typically want to understand. For example, suppose we are
interested in student satisfaction with school and cognitive ability (loosely, how happy and how
smart). Suppose we measure each student's satisfaction with a survey full of items like (my
teachers are...enthusiastic, stupid, putrid, helpful; the school facilities are...well lighted,
crumbling, smelly, inspiring; overall, my school is ... great, boring, fun, interesting, a waste of
time). Responses are summed from strongly disagree to strongly agree; reverse coded, etc. For
cognitive ability, we measure student grade point average. Now let's consider what the
correlation between the two might mean.
Our correlation will be based on two observed measures, one for satisfaction and one for GPA.
Observed measures are represented by square boxes. The observed measures usually don't cause
anything, but rather reflect other causes. For example, part of GPA is

due to cognitive ability. Part of our satisfaction measure is due to affective reactions to school
(what we want to measure). Both observed measures are subject to errors of measurement (e1
and e2). Both are subject to stray causes (e.g., motives). Now the correlation between the two
observed measures could be due to a simple correlation between the things of interest, one
causing the other (a to b or b to a), a third variable, or shared errors. This is the difficulty we face
when trying to interpret r as a measure of influence.