Chapter 06.

03
Linear Regression

After reading this chapter, you should be able to

1. define regression,
2. use several minimizing of residual criteria to choose the right criterion,
3. derive the constants of a linear regression model based on least squares method
criterion,
4. use in examples, the derived formulas for the constants of a linear regression model,
and
5. prove that the constants of the linear regression model are unique and correspond to
a minimum.
Linear regression is the most popular regression model. In this model, we wish to
predict response to n data points ( x1 , y1 ), ( x 2 , y 2 ),......, ( x n , y n ) by a regression model given
by
(1)
y = a0 + a1 x
where a 0 and a1 are the constants of the regression model.
A measure of goodness of fit, that is, how well a0 + a1 x predicts the response variable
y is the magnitude of the residual i at each of the n data points.
(2)
Ei = y i (a 0 + a1 xi )
Ideally, if all the residuals i are zero, one may have found an equation in which all
the points lie on the model. Thus, minimization of the residual is an objective of obtaining
regression coefficients.
The most popular method to minimize the residual is the least squares methods,
where the estimates of the constants of the models are chosen such that the sum of the
squared residuals is minimized, that is minimize

E
i =1

Why minimize the sum of the square of the residuals? Why not, for instance,
minimize the sum of the residual errors or the sum of the absolute values of the residuals?
Alternatively, constants of the model can be chosen such that the average residual is zero
without making individual residuals small. Will any of these criteria yield unbiased
06.03.1

06.03.2

Chapter 06.03

parameters with the smallest variance? All of these questions will be answered below. Look
at the data in Table 1.
Table 1 Data points.

x
2.0
3.0
2.0
3.0

y
4.0
6.0
6.0
8.0

To explain this data by a straight line regression model,

y = a 0 + a1 x
and using minimizing

E as a criteria to find a
i

i =1

y = 4x 4

(3)

and a1 , we find that for (Figure 1)

(4)

10

y = 4x 4

6
4
2
0
0

Figure 1 Regression curve y = 4 x 4 for y vs. x data.

the sum of the residuals,

E
i =1

= 0 as shown in the Table 2.

Table 2 The residuals at each data point for regression model y = 4 x 4 .

y
y predicted = y y predicted
x
2.0
3.0
2.0
3.0

4.0
6.0
6.0
8.0

4.0
8.0
4.0
8.0

0.0
-2.0
2.0
0.0
4

i =1

=0

Linear Regression

06.03.3

So does this give us the smallest error? It does as

E
i =1

= 0 . But it does not give

unique values for the parameters of the model. A straight-line of the model
y= 6
also makes

E
i =1

(5)

= 0 as shown in the Table 3.

Table 3 The residuals at each data point for regression model y = 6

y
y predicted = y y predicted
x
2.0
3.0
2.0
3.0

4.0
6.0
6.0
8.0

6.0
6.0
6.0
6.0

-2.0
0.0
0.0
2.0
4

E
i =1

=0

y=6

7
5
3
1.5

2.5

3.5

Figure 2 Regression curve y = 6 for y vs. x data.

Since this criterion does not give a unique regression model, it cannot be used for
finding the regression coefficients. Let us see why we cannot use this criterion for any
general data. We want to minimize
n

i =1

i =1

Ei = ( yi a0 a1 xi )

(6)

Differentiating Equation (6) with respect to a 0 and a1 , we get

n

Ei
i =1

a0

= 1 = n
i =1

(7)

06.03.4

Chapter 06.03
n

Ei

= xi = n x

(8)
a1
i =1
Putting these equations to zero, give n = 0 but that is not possible. Therefore, unique values
of a 0 and a1 do not exist.
i =1

You may think that the reason the minimization criterion

E
i =1

negative residuals cancel with positive residuals. So is minimizing

does not work is that

n

E
i =1

look at the data given in the Table 2 for equation y = 4 x 4 . It makes

E
i =1

in the following table.

better? Let us
i

= 4 as shown

Table 4 The absolute residuals at each data point when employing y = 4 x 4 .

y
y predicted
= y y predicted
x
2.0
3.0
2.0
3.0

4.0
6.0
6.0
8.0

4.0
8.0
4.0
8.0

0.0
2.0
2.0
0.0
4

i =1

The value of

E
i =1

=4

= 4 also exists for the straight line model y = 6 . No other straight line
4

model for this data has Ei < 4 . Again, we find the regression coefficients are not unique,
i =1

and hence this criterion also cannot be used for finding the regression model.
Let us use the least squares criterion where we minimize
n

S r = Ei = ( yi a0 a1 xi )
2

i =1

(9)

i =1

S r is called the sum of the square of the residuals.

To find a 0 and a1 , we minimize S r with respect to a 0 and a1 .

giving

n
S r
= 2 ( y i a 0 a1 xi )( 1) = 0
a 0
i =1
n
S r
= 2 ( y i a 0 a1 xi )( xi ) = 0
a1
i =1
n

i =1

i =1

i =1

yi + a0 + a1 xi = 0

(10)
(11)

(12)

Linear Regression

06.03.5

i =1

i =1

i =1

y i x i + a 0 x i + a1 x i2 = 0
n

Noting that

i =1

(13)

= a 0 + a 0 + . . . + a 0 = na 0

i =1

i =1

na 0 + a1 xi = y i

(14)

i =1

i =1

i =1

a 0 x i + a1 x i2 = x i y i

(15)

( xn , y n )

(xi , yi )

Ei = yi a0 a1 xi

( x2 , y 2 )
(x3 , y3 )
y = a 0 + a1 x

(x1 , y1 )

Figure 3 Linear regression of y vs. x data showing residuals and square of residual at a
typical point, xi .
Solving the above Equations (14) and (15) gives
n

a1 =

a0 =
Redefining

n x i y i x i y i
i =1

i =1

i =1

n x i2 x i
i =1
i =1
n

i =1

i =1

i =1

i =1
2

xi2 y i xi xi y i

n x i2 x i
i =1
i =1
n

(16)

(17)

06.03.6

Chapter 06.03
n

_ _

S xy = x i y i n x y

(18)

i =1

_2

S xx = x n x
2
i

i =1

(19)

x=

x
i =1

(20)

n
n

y=

y
i =1

(21)

n
we can rewrite
S xy
a1 =
S xx
_

(22)
_

a 0 = y a1 x

(23)

Example 1
The torque T needed to turn the torsional spring of a mousetrap through an angle, is given
below
Table 5 Torque versus angle for a torsion spring.
Torque, T
Angle,
Nm
Radians
0.698132
0.188224
0.959931
0.209138
1.134464
0.230052
1.570796
0.250965
1.919862
0.313707
Find the constants k1 and k 2 of the regression model
T = k 1 + k 2
Solution
Table 6 shows the summations needed for the calculation of the constants of the regression
model.
Table 6 Tabulation of data for calculation of needed summations.
i

T
T
2
2
1
radians
Nm
Nm
radians
1
2
0.698132 0.188224 4.87388 10
1.31405 10 1
3
0.959931 0.209138 9.21468 10 1 2.00758 10 1
4
1.134464 0.230052 1.2870
2.60986 10 1

Linear Regression

06.03.7

5
6
5

1.570796 0.250965 2.4674

1.919862 0.313707 3.6859

3.94215 10 1
6.02274 10 1

6.2831

1.5896

1.1921

8.8491

i =1

n=5

n i Ti i Ti

k2 =

i =1

i =1

i =1

n i2 i
i =1
i =1
5

5(1.5896) (6.2831)(1.1921)
5(8.8491) (6.2831) 2
= 9.6091 10 2 N - m/rad
=

T=

T
i =1

n
1.1921
=
5
= 2.3842 10 1 N-m
5

i =1

n
6.2831
=
5
= 1.2566 radians
_

k1 = T k 2
= 2.3842 10 1 (9.6091 10 2 )(1.2566)
= 1.1767 10 1 N - m

06.03.8

Chapter 06.03

Figure 4 Linear regression of torque vs. angle data

Example 2
To find the longitudinal modulus of a composite material, the following data, as given in
Table 7, is collected.
Table 7 Stress vs. strain data for a composite material.
Strain Stress
(%)
( MPa )
0
0
0.183 306
0.36
612
0.5324 917
0.702 1223
0.867 1529
1.0244 1835
1.1774 2140
1.329 2446
1.479 2752
1.5
2767
1.56
2896
Find the longitudinal modulus E using the regression model.

Linear Regression

06.03.9

= E

(24)

Solution
Rewriting data from Table 7, stresses versus strain data in Table 8
Table 8 Stress vs strain data for a composite in SI system of units
Strain
Stress
( m/m )
( Pa )
0.0000
0.0000
3
1.8300 10
3.0600 10 8
3.6000 10 3 6.1200 10 8
5.3240 10 3 9.1700 10 8
7.0200 10 3 1.2230 10 9
8.6700 10 3 1.5290 10 9
1.0244 10 2 1.8350 10 9
1.1774 10 2 2.1400 10 9
1.3290 10 2 2.4460 10 9
1.4790 10 2 2.7520 10 9
1.5000 10 2 2.7670 10 9
1.5600 10 2 2.8960 10 9
Applying the least square method, the residuals i at each data point is
i = i E i
The sum of square of the residuals is
n

S r = i2
i =1
n

= ( i E i )

i =1

Again, to find the constant E , we need to minimize S r by differentiating with respect to E

and then equating to zero
n
dS r
= 2( i E i )( i ) = 0
dE i =1
From there, we obtain
n

E=

i =1
n

i =1

(25)

Note, Equation (25) only so far has shown that it corresponds to a local minimum or
maximum. Can you show that it corresponds to an absolute minimum.
The summations used in Equation (25) are given in the Table 9.

06.03.10

Chapter 06.03
Table 9 Tabulation for Example 2 for needed summations

i
1
2
3
4
5
6
7
8
9
10
11
12

0.0000
1.8300 10 3
3.6000 10 3
5.3240 10 3
7.0200 10 3
8.6700 10 3
1.0244 10 2
1.1774 10 2
1.3290 10 2
1.4790 10 2
1.5000 10 2
1.5600 10 2

0.0000
3.0600 10 8
6.1200 10 8
9.1700 10 8
1.2230 10 9
1.5290 10 9
1.8350 10 9
2.1400 10 9
2.4460 10 9
2.7520 10 9
2.7670 10 9
2.8960 10 9

12

i =1

n = 12
12

i =1
12

2
i

= 1.2764 10 3

i =1

= 2.3337 10 8

12

E=

i =1
12

i i

i =1

2.3337 10 8
1.2764 10 3
= 182.84 GPa
=

2
0.0000
3.3489 10 6

1.2960 10 5
2.8345 10 5
4.9280 10 5
7.5169 10 5
1.0494 10 4
1.3863 10 4
1.7662 10 4
2.1874 10 4
2.2500 10 4
2.4336 10 4

0.0000
5.5998 10 5
2.2032 10 6
4.8821 10 6
8.5855 10 6
1.3256 10 7
1.8798 10 7
2.5196 10 7
3.2507 10 7
4.0702 10 7
4.1505 10 7
4.5178 10 7

1.2764 10 3

2.3337 10 8

Linear Regression

06.03.11

Figure 5 Linear regression model of stress vs. strain for a composite material.
QUESTION:
Given n data pairs, ( x1 , y1 ), , ( x n , y n ) , do the values of the two constants a0 and a1 in the
least squares straight-line regression model y = a 0 + a1 x correspond to the absolute minimum
of the sum of the squares of the residuals? Are these constants of regression unique?
ANSWER:
Given n data pairs ( x1 , y1 ), , ( x n , y n ) , the best fit for the straight-line regression model
(A.1)
y = a 0 + a1 x
is found by the method of least squares. Starting with the sum of the squares of the residuals
Sr
n

S r = ( y i a 0 a1 xi )

(A.2)

i =1

and using
S r
=0
a 0
S r
=0
a1
gives two simultaneous linear equations whose solution is

(A.3)
(A.4)

06.03.12

Chapter 06.03
n

a1 =

i =1

i =1

i =1

n x xi
i =1
i =1
n

2
i

a0 =

n xi y i xi y i

(A.5a)

xi2 yi xi xi yi

(A.5b)

n xi2 xi
i =1
i =1
But do these values of a0 and a1 give the absolute minimum of value of S r (Equation (A.2))?
The first derivative analysis only tells us that these values give a local minima or maxima of
S r , and not whether they give an absolute minimum or maximum. So, we still need to figure
out if they correspond to an absolute minimum.
i =1

i =1
n

i =1

i =1
2

We need to first conduct a second derivative test to find out whether the point (a 0 , a1 ) from
Equation (A.5) gives a local minimum or local maximum of S r . Only then can we proceed
to show if this local minimum (or maximum) also corresponds to the absolute minimum (or
maximum).
What is the second derivative test for a local minimum of a function of two variables?
If you have a function f ( x, y ) and we found a critical point (a, b ) from the first derivative
test, then (a, b ) is a minimum point if
2

2 f 2 f 2 f
> 0 , and

x 2 y 2 xy
2 f
2 f
OR
>
0
>0
x 2
y 2
From Equation (A.2)
n
S r
= 2( y i a 0 a1 xi )(1)
a 0 i =1

(A.6)
(A.7)

(A.8)

= 2 ( y i a 0 a1 xi )
i =1

S r
= 2 y i a 0 a1 xi ( xi )
a1 i =1
n

(
n

= 2 x i y i a 0 x i a x

then

i =1

n
2Sr
=

2
1 = 2n

a 02
i =1

2
1 i

(A.9)

(A.10)

Linear Regression

06.03.13

n
2Sr
=
2
xi2

2
a1
i =1

(A.11)

n
2Sr
= 2 x i
a 0 a1
i =1

(A.12)

So, we satisfy condition (A.7) because from Equation (A.10) we see that 2n is a positive
n

number. Although not required, from Equation (A.11) we see that 2 xi2 is also a positive
i =1

number as assuming that all x data points are NOT zero is reasonable.
Is the other condition (Equation (A.6)) for S r being a minimum met? Yes, we can show
(proof not given that the term is positive)
2

2 Sr 2 Sr 2 Sr
n 2 n

(
)
=
2
n

2 x i 2 x i
a02 a12 a0 a1
i =1 i =1
n 2 n 2
= 4 n xi xi
i =1
i =1

= 4 ( x i x j ) 2 > 0

(A.13)

i =1
i< j

So the values of a0 and a1 that we have in Equation (A.5) do correspond to a local minimum
of S r . But, is this local minimum also an absolute minimum. Yes, as given by Equation
(A.5), the first derivatives of S r are zero at only one point. This observation also makes the
straight-line regression model based on least squares to be unique.
As a side note, the denominator in Equations (A.5) is nonzero as shown by Equation (A.13).
This shows that the values of a0 and a1 are finite.
LINEAR REGRESSION
Topic
Linear Regression
Summary Textbook notes of Linear Regression
Major
General Engineering
Authors
Egwu Kalu, Autar Kaw, Cuong Nguyen
Date
August 13, 2012
Web Site http://numericalmethods.eng.usf.edu