CORRELATION and REGRESSION

Earlier we discussed the relationship between two or more

variables that were categorical. Here we discuss the
relationship between two or more variables that are
quantitative. (This is an introductory course so we will
stick to two variables.)

Correlation means to study the relationship between

variables, while regression involves making predictions of
one variable based on other variables.

Response variable: a variable that measures an outcome

Explanatory variable: a variable the may explain or

influence a response variable

There can be many response and explanatory variables and

in fact this happens quite often. However, we will only
worry about one of each and call the response variable y
and the explanatory variable x.

Sometimes we are just interested in studying the

relationship between variables in this case no distinction
needs to be made between the two.

As an example a person interested in health might look for

relationships between many measures of health such as
BMI, blood pressure, cholesterol level, etc. In this case no
distinction needs to be made.
If you wish to predict one (or more) variable from another
(or several others) the one(s) that are to be predicted are
response variable(s) and the one(s) the prediction is based
on are the explanatory variable(s).

As an example you might want to use such things as BMI,

blood pressure, cholesterol level and other variables to
predict the life expectancy of a person. In this case the
BMI, blood pressure, cholesterol level and so forth are
explanatory variables and the life expectancy is the
response variable.

Scatterplot: A scatterplot shows the relationship between

two quantitative variables measured on the same
individuals. Each individual is one dot. If there is an
explanatory variable put it on the x-axis.

Example: What do you think of ranking state school

systems by average SAT scores? Is a 559 state SAT math
average better than a 548? At first you may think so but
there is another variable that might affect SAT math
average other than the state’s school system. This other
variable is the percent of student’s that actually take the
SAT.

Below is a scatterplot of the 50 states and DC. On the x-

axis is the percent taking the SAT and on the y-axis is the
SAT math average. We think that the percent taking might
explain the average so it goes on the x-axis.
In examining a scatterplot you should look for the overall
pattern (direction, form, strength) and any outliers.

What do you see in the scatterplot with SAT math

averages?
650

600
SAT math average

550

500

450

400
0 20 40 60 80 100
Percent in a state taking SAT

This scatterplot has negative relationship. That means that

as x increases, y tends to decrease. A positive relationship
will have the property that as x increases, y tends to also
increase. Scatterplots do not have to have a positive or a
negative relationship, they can certainly be neither.

In our SAT example you could actually see which states are
doing the best. It is not the ones with the highest SAT math
averages, but rather the states that have their dots towards
the upper part of the general pattern. (The states that have
their dots towards the bottom are not doing as well.)

The data for the scatterplot was for 2002. Colorado had
28% take the SAT and a SAT math average of 548.
Alabama had 9% take the SAT and a math average of 559.
Which state do you think is better?

Recall the Palm Beach scatterplot:

This scatterplot shows a positive relationship that is

basically linear with an extreme outlier!

Recall two-way tables. We saw this one before.

Instructor Assistant Associate Full Total

Professor Professor Professor
Female 61 565 502 476 1604
Male 39 618 683 1383 2723
Total 100 1183 1185 1859 4327

Notice that it is like a scatterplot but instead of putting

4327 dots which can only be in 8 places we tell how many
dots are in each of the 8 places. So a two-way table is
almost like the categorical data version of a scatterplot.

Here are a couple of scatterplots that show a strong

relationship but it is not linear.
Here is an example of a relationship that at places is
positive and at others is negative and is quite strong. This
graph is not a scatterplot, but was obtained from a
scatterplot. (It shows the curve of best fit.)

Sometimes we want to add categorical variables to a

scatterplot. For example in the SAT math example,
perhaps you might want to color the southern states a
different color. If you could obtain data from say 60 years
ago guess where the southern states would be. Where do
you think they are in 2002?

MEASURING THE STRENGTH OF A LINEAR

RELATIONSHIP

The population number is (rho), which we won’t

calculate, the sample number is r.

LINE OF BEST FIT

Sometimes it makes sense to find the Prediction Line =
Regression Line = Line of Best Fit = Least Square’s Line =
Least Square’s Regression Line. (They are all the same
thing.) This is a line through the scatterplot that minimizes
the sum of the squares of how far vertically the points are
from the line. The following give the equation for the line
of best fit for sample data:

FACTS AND CAUTIONS ABOUT CORRELATION

AND REGRESSION

If you switch x and y the correlation stays the same. You

can tell by looking at the formula for r.

r does not change if we change the units of measurement

on x and/or y. This is because r uses standardized values.
Standardized means how many standard deviations from
the mean. If you measure your weight in both pounds and
find that you are 1.43 standard deviations above the mean
and then convert to kilograms you will still be 1.43
standard deviations above the mean.

If r is positive there is a positive association between x and

y. If r is negative there is a negative association between x
and y. You can tell by looking at the formula for r. If there
is a positive relationship then bigger than average x’s will
tend to have bigger than average y’s and smaller than
average x’s will tend to have smaller than average y’s. This
will tend to make
either the product of two positives or the
product of two negatives, which are both positive. In the
formula for r these are added up. If you add up numbers
that tend to be positive you will get a positive. On the other
hand, if there is a negative relationship then bigger than
average x’s will tend to have smaller than average y’s and
smaller than average x’s will tend to have larger than
average y’s. This will tend to make
the product of a negative and a positive which
is negative. In the formula for r these are added up. If you
add up numbers that tend to be negative you will get a
negative.

r is always between -1 and 1(not obvious, but can be

proven). r near 0 indicates a very weak linear relationship.
As r moves away from 0 the linear relationship is stronger.
At r=1 and r=-1 you have perfect linear relationships. We
can see some examples of scatterplots with different r’s by
following the link at the end of these slides.

r measures only the strength of a linear relationship. It will

not describe how strong other relationships are. You
should plot your data before finding r. You may have a
case with r=.5, a somewhat weak linear relationship, but
the relationship is extremely strong, just not linear. Again
we can see examples with the link at the end of these slides.

Like the mean and standard deviation r is not resistant to

outliers. These are all strongly affected by a few extreme
observations. Again we can see examples with the link at
the end of these slides.

If you switch x and y you get different regression lines.

You can see why this is so by recalling that the regression
line minimizes the sum of the squares of the vertical errors.
If you switch x and y then instead of vertical errors it will
be horizontal errors.

Because , a change in 1 standard deviation of x results

in a change of r standard deviations of y. Recall that slope
is rise over run and so it tells us the general change in y
relative to x. Also recall that how many standard
deviations from the mean (standardized scores) allows for
universal comparisons. If we had x=height and y=weight
measured in inches and pounds then the slope would give
us the how many pounds per inch on average. But we
could also conclude that an increase in one standard
deviation in height would result in r standard deviations
increase in weight.

tells the percent of differences in y attributed to the

regression line on x. There may be many factors that
explain the differences in y, the one we are studying is x. If
you attempt to find other factors that explain the
differences in y, you may very well have your percents
greatly exceed 100%. This is because most likely the
factors will overlap. y varies for two reasons. There is
scatter and because x changes. measures the variation
due to x changing divided by the total variation. This fact
is not obvious but can be proven.

You should not find the line of best fit unless the data
shows a linear pattern. You can do the calculations no
matter, but it may be nonsense. Again we can see
examples with the link at the end of these slides.

The regression line, just like r, is not resistant to outliers.

Again we can see examples with the link at the end of these
slides.

Extrapolation is the use of the regression line far outside

the range of values obtained for x. These predictions are
almost never trustworthy. As an example, suppose you
collect data from an elementary school for students. You
record the ages and the heights and find the line of best fit.
Then you try to predict the height of someone that you
know is 80 years old. What would happen? The thing to
remember is that predictions are only somewhat reliable
near where your data was collected.

Association does not imply causation. A high correlation

does not mean that x causes y, rather that x explains y.
Recall the example we mentioned about ice cream sales
and sunburns. Suppose we measured ice cream sales and
sunburns each day in some mountain resort. What would
we find? Does this mean that the ice cream is causing the
sunburns? Many times people try to claim a high
correlation means causation. As you become more
statistically literate, you will not be fooled by such people.

Often even when there is causation it may not be the entire

explanation for the correlation. As an example suppose
someone is interested in genetic causation of BMI’s in
mothers and daughters. Data is collected and it is found
that for the correlation of BMI’s for mothers and daughters
that r=0.55. This suggests a genetic causation. However,
there is also an environmental factor. Daughters learn
many habits from their mothers that will affect their BMI,
such as exercise and eating habits. Even though there is
some genetic causation the r=0.55 not entirely due to
genetics, rather it is a mix of genetics and environment. It
is extremely difficult to find out how much genetics is
contributing to the value of r. Note that it is even possible
(although unlikely) that the environment factor is actually
negative and the genetic factor contributes more than 0.55
to the value of r. This would happen for example if the
daughters saw bad habits and their consequences of their
mothers and learned to do the opposite. This may happen
from time to time but these are surely anecdotes.

We could find the true genetic causation by removing the

daughters from their mothers, but people would complain!

It is usually quite difficult to determine how much one

variable directly affects another one when there are many
other variables lurking in the background without doing an
experiment. Experiments divide subjects up randomly and
by doing so eliminate lurking variables. Recall
experiments defeated the lurking variable of the women’s
importance of their health from the studies that suggested
that hormone replacement therapy would cause a reduction
in heartattacks in women at menopause.

However, experiments are not always possible.

Can we use statistics to determine some of the questions

below? What is the genetic effect of mother’s BMI on
daughter’s BMI? Would stricter gun control laws reduce or
increase violent crime? Has increased free trade widened
the gap between the incomes of the less educated and more
educated? In all these cases experiments are not practical.
Statistics can try to answer these questions, but it can very
difficult or even impossible. Trying to pinpoint cause and
effect in settings involving complex relations between
many variables is nearly impossible if experiments can’t be
done. Still, people that have specific points of view will try
to use statistics to further their agenda. Do not fall prey to
their shaky arguments.

Next we will give two examples in which no experiment

have been done and give potential lurking variables that
could explain the association and hence mean there is no
causation.

Does smoking cause lung cancer? No experiments have

been done on humans. There is a strong association
between smoking and lung cancer. That does not mean that
smoking causes lung cancer, recall there is a strong
association between ice cream sales and sunburns, and ice
cream sales do not cause sunburns! Perhaps there is a
lurking variable such as a gene that does two things. It
causes people to tend to be addicted to nicotine and it tends
to cause tumors in the lungs of people. If this were the
case, it would explain why there is such a strong
association between smoking and lung cancer. It would be
the gene that caused cancer and not the smoking itself.
Tobacco executives not too long ago claimed in front of
congress that there was no evidence that smoking caused
lung cancer. People have pretty much ruled the possibility
of such a lurking variable. People who smoke more are
more likely to get lung cancer. People who stop smoking
reduce their risk. Lung cancer increased dramatically as
smoking became more popular (with a lag of about 30
years, since lung cancer develops after years of smoking).
Lung cancer was rare among women until smoking became
popular among women (again with a lag of about 30 years).
Experiments on animals show that the tars from smoking
do cause cancer. With all these facts it is unlikely such a
genetic factor exists. If it did, this gene started showing up
in humans at the same time that smoking became popular!
Perhaps there is another lurking variable we haven’t
thought of. Medical authorities don’t hesitate to say that
smoking causes lung cancer. It is written on packs of
cigarettes! They have very good evidence of this. But the
evidence would be better if experiments were done on
people, but this almost surely will never happen.
Does being educated cause people to be wealthier? There
is a strong association between education level and wealth.
But no experiments have been done. We would have to
divide people up at random and educate some and not
others! Perhaps there is a lurking variable such as how
motivated a person is. Being motivated would cause
people to be wealthy and also get an education. If a
motivated person did not have the chance for an education
perhaps they would put their effort towards becoming
wealthy? Do you think that this lurking variable
(motivation level) explains all the association between
education and wealth? Do you think that it explains part of
it?

We have seen that lurking variables can make it appear that

x has a strong direct affect on y when it doesn’t. As another
example consider TV sets per 1000 people and life
expectancy of the countries of the world.
 tvs per 1000 and life expectency of 27 countries
picked at random

85

80

75

70
life expectancy

65

60

55

50

45

40
0 200 400 600 800 1000
tvs per 1000

From the graph it may look like we could cause an increase

in life expectancies of poorer countries by shipping them
TVs. Do you think this would work?

However, lurking variables can also hide an effect or at

least make a causation appear not as strong as it really is.
Consider data on a household. Degree days per month and
amount of electricity used in a month are given (assume the
house is heating with electricity and is in the high
mountains of Colorado so no AC is needed in the summer).
A degree day is the number of degrees that the mean
temperature for that day is below 65° F. Suppose that data
is keep for several years but that just over a year ago the
household installed a few solar panels to produce
electricity. Suppose the solar panels were not mentioned
and whether or not there were solar panels was a lurking
variable. Here is the scatterplot.

300

250
montly electricity bill

200

150

100

50

0
0 200 400 600 800 1000 1200
degree days per month

The relationship is apparent and not surprising. However if

we use different dots to show the months in which the solar
panels were in use we get the following:
 300

250
montly electricity bill

200

Without Solar Panels

150
With Solar Panels

100

50

0
0 200 400 600 800 1000 1200
degree days per month

Notice the association is much stronger within each

category than when looked at overall. Here is another
reason to include categories in a scatterplot (recall the
southern states example earlier).

Below is a scatterplot where each dot represents a state.

We are comparing SAT Math averages with SAT Verbal
averages. Notice a very strong relationship. What do you
think would happen if instead of 50 states we had 50
individual students picked at random? Do you think the
relationship would be stronger or weaker or the same?
Another way of asking this is, “do you think there would be
more or less or the same amount of scatter?”
 640
SAT V

540

440
440 540 640
SAT M

It is important to remember that if your data is averages

you should not apply correlation and regression to
individuals with your data. The correlation for averages is
much higher than individuals and any prediction for
individuals based on averages will not be as trustworthy.

There are many other types of regression and correlation

besides linear. Examples are quadratic, square root, cubic,
exponential, logarithmic, and logistic, just to name a few.
There are hundreds of possible forms. Below are the
graphs of some common forms.

Quadratic .
Exponential .

Logarithmic .

Logistic

There is a neat website to play with scatterplots and

regression lines, it is

https://www.stat.berkeley.edu/~stark/Java/Html/
Correlation.htm
or
http://digitalfirst.bfwpub.com/stats_applet/
stats_applet_5_correg.html