Correlation and
Regression
Cal State Northridge
427
Ainsworth
Major Points - Correlation
Questions answered by correlation
Scatterplots
An example
The correlation coefficient
Other kinds of correlations
Factors affecting correlations
Testing for significance
The Question
Are two variables related?
Does one increase as the other increases?
e. g. skills and income
Does one decrease as the other increases?
e. g. health problems and nutrition
How can we get a numerical measure of
the degree of relationship?
Scatterplots
AKA scatter diagram or scattergram.
Graphically depicts the relationship
between two variables in two dimensional
space.
Direct Relationship
Scatterplot:Video Games and Alcohol Consumption
20
Average Number of Alcoholic Drinks
18
16
14
Per Week
12
10
8
6
4
2
0
0 5 10 15 20 25
Average Hours of Video Games Per Week
Inverse Relationship
Scatterplot: Video Games and Test Score
100
90
80
70
Exam Score
60
50
40
30
20
10
0
0 5 10 15 20
Average Hours of Video Games Per Week
An Example
Does smoking cigarettes increase systolic
blood pressure?
Plotting number of cigarettes smoked per
day against systolic blood pressure
Fairly moderate relationship
Relationship is positive
Trend?
170
160
150
140
130
120
SYSTOLIC
110
100
0 10 20 30
SMOKING
Smoking and BP
Note relationship is moderate, but real.
Why do we care about relationship?
What would conclude if there were no
relationship?
What if the relationship were near perfect?
What if the relationship were negative?
Heart Disease and Cigarettes
Data on heart disease and cigarette
smoking in 21 developed countries
(Landwehr and Watkins, 1987)
Data have been rounded for computational
convenience.
The results were not affected.
Country Cigarettes CHD
The Data
1 11 26
2 9 21
3 9 24
4 9 21
5 8 19
6 8 13
7 8 19
Surprisingly, the 8 6 11
9 6 23
U.S. is the first 10 5 15
country on the list- 11 5 13
12 5 4
-the country 13 5 18
with the highest 14 5 12
15 5 3
consumption and 16 4 11
17 4 15
highest mortality. 18 4 6
19 3 13
20 3 4
21 3 14
Scatterplot of Heart Disease
CHD Mortality goes on ordinate (Y axis)
Why?
Cigarette consumption on abscissa (X
axis)
Why?
What does each dot represent?
Best fitting line included for clarity
30
20
10
{X = 6, Y = 11}
0
2 4 6 8 10 12
Cigarette Consumption per Adult per Day
What Does the Scatterplot Show?
As smoking increases, so does coronary
heart disease mortality.
Relationship looks strong
Not all data points on line.
This
gives us “residuals” or “errors of
prediction”
To be discussed later
Correlation
Co-relation
The relationship between two variables
Measured with a correlation coefficient
Most popularly seen correlation
coefficient: Pearson Product-Moment
Correlation
Types of Correlation
Positive correlation
High values of X tend to be associated with
high values of Y.
As X increases, Y increases
Negative correlation
High values of X tend to be associated with
low values of Y.
As X increases, Y decreases
No correlation
No consistent tendency for values on Y to
increase or decrease as X increases
Correlation Coefficient
A measure of degree of relationship.
Between 1 and -1
Sign refers to direction.
Based on covariance
Measure of degree to which large scores on
X go with large scores on Y, and small scores
on X go with small scores on Y
Think of it as variance, but with 2 variables
instead of 1 (What does that mean??)
18
Covariance
Remember that variance is:
( X X ) 2
( X X )( X X )
VarX
N 1 N 1
The formula for co-variance is:
( X X )(Y Y )
Cov XY
N 1
How this works, and why?
When would covXY be large and positive?
Large and negative?
Country X (Cig.) Y (CHD) (X X ) (Y Y ) ( X X ) * (Y Y )
1 11 26 5.05 11.48 57.97
2 9 21 3.05 6.48 19.76
3 9 24 3.05 9.48 28.91
4 9 21 3.05 6.48 19.76
5 8 19 2.05 4.48 9.18
6 8 13 2.05 -1.52 -3.12
7 8 19 2.05 4.48 9.18
8 6 11 0.05 -3.52 -0.18
9 6 23 0.05 8.48 0.42
Example
10 5 15 -0.95 0.48 -0.46
11 5 13 -0.95 -1.52 1.44
12 5 4 -0.95 -10.52 9.99
13 5 18 -0.95 3.48 -3.31
14 5 12 -0.95 -2.52 2.39
15 5 3 -0.95 -11.52 10.94
16 4 11 -1.95 -3.52 6.86
17 4 15 -1.95 0.48 -0.94
18 4 6 -1.95 -8.52 16.61
19 3 13 -2.95 -1.52 4.48
20 3 4 -2.95 -10.52 31.03
21 3 14 -2.95 -0.52 1.53
Mean 5.95 14.52
SD 2.33 6.69
Sum 222.44
Example
21
( X X )(Y Y ) 222.44
Covcig .&CHD 11.12
N 1 21 1
What the heck is a covariance?
I thought we were talking about
correlation?
Correlation Coefficient
Pearson’s Product Moment Correlation
Symbolized by r
Covariance ÷ (product of the 2 SDs)
Cov XY
r
s X sY
Correlation is a standardized
covariance
Calculation for Example
CovXY = 11.12
sX = 2.33
sY = 6.69
cov XY 11.12 11.12
r .713
s X sY (2.33)(6.69) 15.59
Example
Correlation = .713
Sign is positive
Why?
If sign were negative
What would it mean?
Would not alter the degree of relationship.
Other calculations
25
Z-score method
r
z z x y
N 1
Computational (Raw Score) Method
N XY X Y
r
N X 2 ( X )2 N Y 2 ( Y )2
Other Kinds of Correlation
Spearman Rank-Order Correlation
Coefficient (rsp)
used with 2 ranked/ordinal variables
uses the same Pearson formula
Attractiveness Symmetry
3 2
4 6
1 1
2 3
5 4
6 5 26
rsp = 0.77
Other Kinds of Correlation
Point biserial correlation coefficient (rpb)
used with one continuous scale and one
nominal or ordinal or dichotomous scale.
uses the same Pearson formula
Attractiveness Date?
3 0
4 0
1 1
2 1
5 1
6 0
rpb = -0.49 27
Other Kinds of Correlation
Phi coefficient ()
used with two dichotomous scales.
uses the same Pearson formula
Attractiveness Date?
0 0
1 0
1 1
1 1
0 0
1 1
= 0.71 28
Factors Affecting r
Range restrictions
Looking at only a small portion of the total
scatter plot (looking at a smaller portion of
the scores’ variability) decreases r.
Reducing variability reduces r
Nonlinearity
The Pearson r (and its relatives) measure the
degree of linear relationship between two
variables
If a strong non-linear relationship exists, r will
provide a low, or at least inaccurate measure
of the true relationship.
Factors Affecting r
Heterogeneous subsamples
Everyday examples (e.g. height and weight
using both men and women)
Outliers
Overestimate Correlation
Underestimate Correlation
Countries With Low Consumptions
Data With Restricted Range
Truncated at 5 Cigarettes Per Day
20
18
16
CHD Mortality per 10,000
14
12
10
4
2
2.5 3.0 3.5 4.0 4.5 5.0 5.5
Cigarette Consumption per Adult per Day
Truncation
32
Non-linearity
33
Heterogenous samples
34
Outliers
35
Testing Correlations
36
So you have a correlation. Now what?
In terms of magnitude, how big is big?
Smallcorrelations in large samples are “big.”
Large correlations in small samples aren’t
always “big.”
Depends upon the magnitude of the
correlation coefficient
AND
The size of your sample.
Testing r
Population parameter =
Null hypothesis H0: = 0
Test
of linear independence
What would a true null mean here?
What would a false null mean here?
Alternative hypothesis (H1) 0
Two-tailed
Tables of Significance
We can convert r to t and test for
significance:
N 2
tr
1 r 2
Where DF = N-2
Tables of Significance
In our example r was .71
N-2 = 21 – 2 = 19
N 2 19 19
tr .71* .71* 6.90
1 r 2
1 .712
.4959
T-crit (19) = 2.09
Since 6.90 is larger than 2.09 reject = 0.
Computer Printout
Printout gives test of significance.
Correlations
CIGARET CHD
CIGARET Pears on Correlation 1 .713**
Sig. (2-tailed) . .000
N 21 21
CHD Pears on Correlation .713** 1
Sig. (2-tailed) .000 .
N 21 21
**. Correlation is significant at the 0.01 level (2-tailed).
Regression
What is regression?
42
How do we predict one variable from
another?
How does one variable change as the
other changes?
Influence
Linear Regression
43
A technique we use to predict the most
likely score on one variable from those
on another variable
Uses the nature of the relationship (i.e.
correlation) between two variables to
enhance your prediction
Linear Regression: Parts
44
Y - the variables you are predicting
i.e. dependent variable
X - the variables you are using to predict
i.e. independent variable
Ŷ - your predictions (also known as Y’)
Why Do We Care?
45
We may want to make a prediction.
More likely, we want to understand the
relationship.
How fast does CHD mortality rise with a
one unit increase in smoking?
Note: we speak about predicting, but
often don’t actually predict.
An Example
46
Cigarettes and CHD Mortality again
Data repeated on next slide
We want to predict level of CHD
mortality in a country averaging 10
cigarettes per day.
Country Cigarettes CHD
1 11 26
47
The Data 2
3
9
9
21
24
4 9 21
5 8 19
Based on the data we have 6 8 13
what would we predict the 7
8
8
6
19
11
rate of CHD be in a country 9 6 23
10 5 15
that smoked 10 cigarettes on 11 5 13
12 5 4
average? 13 5 18
14 5 12
First, we need to establish a 15 5 3
4 11
prediction of CHD from 16
17 4 15
smoking… 18
19
4
3
6
13
20 3 4
21 3 14
30
We predict a
20
CHD rate of
about 14
Regression
Line
10
For a country that
smokes 6 C/A/D…
0
2 4 6 8 10 12
Cigarette Consumption per Adult per Day
48
Regression Line
49
Formula
Yˆ bX a
Yˆ
= the predicted value of Y (e.g. CHD
mortality)
X = the predictor variable (e.g. average
cig./adult/country)
Regression Coefficients
50
“Coefficients” are a and b
b = slope
Change in predicted Y for one unit change
in X
a = intercept
value of Yˆ when X = 0
Calculation
51
Slope cov XY sy
b 2 or b r
sX sx
N XY X Y
or b
N X 2 ( X ) 2
Intercept a Y bX
For Our Data
52
CovXY = 11.12
s2X = 2.332 = 5.447
b = 11.12/5.447 = 2.042
a = 14.524 - 2.042*5.952 = 2.32
See SPSS printout on next slide
Answers are not exact due to rounding error and desire to match
SPSS.
SPSS Printout
53
Note:
54
The values we obtained are shown on
printout.
The intercept is the value in the B
column labeled “constant”
The slope is the value in the B column
labeled by name of predictor variable.
Making a Prediction
55
Second, once we know the relationship
we can predict
Yˆ bX a 2.042 X 2.367
Yˆ 2.042*10 2.367 22.787
We predict 22.77 people/10,000 in a
country with an average of 10 C/A/D
will die of CHD
Accuracy of Prediction
Finnish smokers smoke 6 C/A/D
We predict:
Yˆ bX a 2.042 X 2.367
Yˆ 2.042*6 2.367 14.619
They actually have 23 deaths/10,000
Our error (“residual”) =
23 - 14.619 = 8.38
a large error
56
30
CHD Mortality per 10,000 Residual
20
Prediction
10
0
2 4 6 8 10 12
Cigarette Consumption per Adult per Day
57
Residuals
58
When we predict Ŷ for a given X, we will
sometimes be in error.
Y – Ŷ for any X is a an error of estimate
Also known as: a residual
We want to Σ(Y- Ŷ) as small as possible.
BUT, there are infinitely many lines that can do
this.
Just draw ANY line that goes through the
mean of the X and Y values.
Minimize Errors of Estimate… How?
Minimizing Residuals
59
Again, the problem lies with this
definition of the mean:
( X X ) 0
So, how do we get rid of the 0’s?
Square them.
Regression Line:
A Mathematical Definition
The regression line is the line which when
drawn through your data set produces the
smallest value of:
(Y Y )
ˆ 2
Called the Sum of Squared Residual or
SSresidual
Regression line is also called a “least squares
line.” 60
Summarizing Errors of Prediction
61
Residual variance
The variability of predicted values
ˆ
(Yi Yi ) 2
SSresidual
s2
Y Yˆ
N 2 N 2
Standard Error of Estimate
62
Standard error of estimate
Thestandard deviation of predicted
values
ˆ
(Yi Yi ) 2
SSresidual
sY Yˆ
N 2 N 2
A common measure of the accuracy of
our predictions
We want it to be as small as possible.
Country X (Cig.) Y (CHD) Y' (Y - Y') (Y - Y')2
Example
1 11 26 24.829 1.171 1.371
2 9 21 20.745 0.255 0.065
3 9 24 20.745 3.255 10.595
63 4 9 21 20.745 0.255 0.065
5 8 19 18.703 0.297 0.088
(Yi Yˆi ) 2 440.756
6 8 13 18.703 -5.703 32.524 2
s
Y Yˆ
23.198
7 8 19 18.703 0.297 0.088 N 2 21 2
8 6 11 14.619 -3.619 13.097
9 6 23 14.619 8.381 70.241 (Yi Yˆi )2 440.756
10 5 15 12.577 2.423 5.871 sY Yˆ
11 5 13 12.577 0.423 0.179
N 2 21 2
12 5 4 12.577 -8.577 73.565 23.198 4.816
13 5 18 12.577 5.423 29.409
14 5 12 12.577 -0.577 0.333
15 5 3 12.577 -9.577 91.719
16 4 11 10.535 0.465 0.216
17 4 15 10.535 4.465 19.936
18 4 6 10.535 -4.535 20.566
19 3 13 8.493 4.507 20.313
20 3 4 8.493 -4.493 20.187
21 3 14 8.493 5.507 30.327
Mean 5.952 14.524
SD 2.334 6.690
Sum 0.04 440.757
Regression and Z Scores
64
When your data are standardized (linearly
transformed to z-scores), the slope of the
regression line is called β
DO NOT confuse this β with the β
associated with type II errors. They’re
different.
When we have one predictor, r = β
Zy = βZx, since A now equals 0
Partitioning Variability
65
Sums of square deviations
Total
SStotal (Y Y )
2
SSregression (Yˆ Y )
2
Regression
Residual we already covered
SSresidual (Y Yˆ )
2
SStotal = SSregression + SSresidual
Partitioning Variability
66
Degrees of freedom
Total
dftotal =N-1
Regression
dfregression = number of predictors
Residual
dfresidual = dftotal – dfregression
dftotal = dfregression + dfresidual
Partitioning Variability
67
Variance (or Mean Square)
Total Variance
s2total = SStotal/ dftotal
Regression Variance
s2regression = SSregression/ dfregression
Residual Variance
s2residual = SSresidual/ dfresidual
Country X (Cig.) Y (CHD) Y' (Y - Y') (Y - Y')2 (Y' - Ybar) (Y - Ybar)
1 11 26 24.829 1.171 1.371 106.193 131.699
2 9 21 20.745 0.255 0.065 38.701 41.939
3 9 24 20.745 3.255 10.595 38.701 89.795
68
4 9 21 20.745 0.255 0.065 38.701 41.939
5 8 19 18.703 0.297 0.088 17.464 20.035
6 8 13 18.703 -5.703 32.524 17.464 2.323
7 8 19 18.703 0.297 0.088 17.464 20.035
8 6 11 14.619 -3.619 13.097 0.009 12.419
9
10
11
6
5
5
23
15
13
14.619
12.577
12.577
8.381
2.423
0.423
70.241
5.871
0.179
0.009
3.791
3.791
71.843
0.227
2.323
Example
12 5 4 12.577 -8.577 73.565 3.791 110.755
13 5 18 12.577 5.423 29.409 3.791 12.083
14 5 12 12.577 -0.577 0.333 3.791 6.371
15 5 3 12.577 -9.577 91.719 3.791 132.803
16 4 11 10.535 0.465 0.216 15.912 12.419
17 4 15 10.535 4.465 19.936 15.912 0.227
18 4 6 10.535 -4.535 20.566 15.912 72.659
19 3 13 8.493 4.507 20.313 36.373 2.323
20 3 4 8.493 -4.493 20.187 36.373 110.755
21 3 14 8.493 5.507 30.327 36.373 0.275
Mean 5.952 14.524
SD 2.334 6.690
Sum 0.04 440.757 454.307 895.247
Y' = (2.04*X) + 2.37
Example
SSTotal (Y Y ) 895.247; dftotal 21 1 20
2
69
(Y Y ) 454.307; df regression 1 (only 1 predictor)
ˆ 2
SSregression
(Y Y ) 440.757; df residual 20 1 19
ˆ 2
SSresidual
2
(Y Y ) 895.247
s2
44.762
N 1
total
20
(Y Y )
ˆ 2
454.307
s2
regression 454.307
1 1
(Y Y )
ˆ 2
440.757
s2
23.198
N 2
residual
19
2
Note : sresidual sY Yˆ
Coefficient of Determination
70
It is a measure of the percent of
predictable variability
r 2 the correlation squared
or
SS regression
r
2
SSY
The percentage of the total variability in
Y explained by X
2
r for our example
71
r = .713
r 2 = .7132 =.508
SSregression 454.307
r
2
.507
or SSY 895.247
Approximately 50% in variability of
incidence of CHD mortality is associated with
variability in smoking.
Coefficient of Alienation
72
It is defined as 1 - r 2 or
SSresidual
1 r 2
SSY
Example
1 - .508 = .492
SS residual 440.757
1 r
2
.492
SSY 895.247
2
r, SS and sY-Y’
73
r2 * SStotal = SSregression
(1 - r2) * SStotal = SSresidual
We can also use r2 to calculate the
standard error of estimate as:
N 1 20
sY Yˆ s y (1 r )
2
6.690* (.492) 4.816
N 2 19
Testing Overall Model
74
We can test for the overall prediction of
the model by forming the ratio:
2
sregression
2
F statistic
sresidual
If the calculated F value is larger than a
tabled value (F-Table) we have a
significant prediction
Testing Overall Model
75
Example 2
sregression 454.307
2
19.594
sresidual 23.198
F-Table – F critical is found using 2 things
dfregression (numerator) and dfresidual.(demoninator)
F-Table our Fcrit (1,19) = 4.38
19.594 > 4.38, significant overall
Should all sound familiar…
SPSS output
76
Model Summary
Adjus ted Std. Error of
Model R R Square R Square the Es timate
1 .713 a .508 .482 4.81640
a. Predictors : (Constant), CIGARETT
ANOVAb
Sum of
Model Squares df Mean Square F Sig.
1 Regress ion 454.482 1 454.482 19.592 .000 a
Res idual 440.757 19 23.198
Total 895.238 20
a. Predictors : (Constant), CIGARETT
b. Dependent Variable: CHD
Testing Slope and Intercept
77
The regression coefficients can be
tested for significance
Each coefficient divided by it’s
standard error equals a t value that
can also be looked up in a t-table
Each coefficient is tested against 0
Testing the Slope
78
With only 1 predictor, the standard
error for the slope is:
sY Yˆ
seb
sX N 1
For our Example:
4.816 4.816
seb .461
2.334 21 1 10.438
Testing Slope and Intercept
79
These are given in computer printout as
a t test.
Testing
80
The t values in the second from right
column are tests on slope and intercept.
The associated p values are next to
them.
The slope is significantly different from
zero, but not the intercept.
Why do we care?
Testing
81
What does it mean if slope is not
significant?
How does that relate to test on r?
What if the intercept is not significant?
Does significant slope mean we predict
quite well?