Chapter 2

Simple Linear Regression Analysis

The simple linear regression model

We consider the modeling between the dependent and one independent variable. When there is only one
independent variable in the linear regression model, the model is generally termed as simple linear
regression model. When there are more than one independent variables in the model, then the linear model
is termed as the multiple linear regression model.

The linear model

Consider a simple linear regression model
y   0  1 X  

where y is termed as the dependent or study variable and X is termed as independent or explanatory
variable. The terms  0 and 1 are the parameters of the model. The parameter  0 is termed as intercept

term and the parameter 1 is termed as slope parameter. These parameters are usually called as regression

coefficients. The unobservable error component  accounts for the failure of data to lie on the straight line
and represents the difference between the true and observed realization of y . There can be several reasons
for such difference, e.g., the effect of all deleted variables in the model, variables may be qualitative, inherit
randomness in the observations etc. We assume that  is observed as independent and identically
distributed random variable with mean zero and constant variance  2 . Later, we will additionally assume
that  is normally distributed.

The independent variables is viewed as controlled by the experimenter, so it is considered as non-stochastic

whereas y is viewed as a random variable with
E ( y )   0  1 X
and
Var ( y )   2 .
Sometimes X can also be a random variable. In such a case, instead of simple mean and simple variance of
y , we consider the conditional mean of y given X  x as
E ( y | x)   0  1 x

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
1
and the conditional variance of y given X  x as

Var ( y | x)   2 .

When the values of  0 , 1 and  2 are known, the model is completely described. The parameters  0 , 1 and

 2 are generally unknown in practice and  is unobserved. The determination of the statistical model
y   0  1 X   depends on the determination (i.e., estimation ) of  0 , 1 and  2 . In order to know the

values of these parameters, n pairs of observations ( xi , yi )(i  1,..., n) on ( X , y ) are observed/collected and
are used to determine these unknown parameters.

Various methods of estimation can be used to determine the estimates of the parameters. Among them, the
methods of least squares and maximum likelihood are the popular methods of estimation.

Least squares estimation

Suppose a sample of n sets of paired observations ( xi , yi ) (i  1, 2,..., n) are available. These observations
are assumed to satisfy the simple linear regression model and so we can write
yi   0  1 xi   i (i  1, 2,..., n).

The principle of least squares estimates the parameters  0 and 1 by minimizing the sum of squares of
difference between the observations and the line in the scatter diagram. Such an idea is viewed from
different perspectives. When the vertical difference between the observations and the line in the scatter
diagram is considered and its sum of squares is minimized to obtain the estimates of  0 and 1 , the method
is known as direct regression. yi

(xi,

Y   0  1 X

(Xi,

xi
Direct regression

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
2
Alternatively, the sum of squares of difference between the observations and the line in horizontal direction
in the scatter diagram can be minimized to obtain the estimates of  0 and 1 . This is known as reverse (or

inverse) regression method.

yi

Y   0  1 X
(xi, yi)

(Xi, Yi)

xi,

Reverse regression method

Instead of horizontal or vertical errors, if the sum of squares of perpendicular distances between the
observations and the line in the scatter diagram is minimized to obtain the estimates of  0 and 1 , the

method is known as orthogonal regression or major axis regression method.

yi

(xi

Y   0  1 X

(Xi
)

xi
Major axis regression method

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
3
Instead of minimizing the distance, the area can also be minimized. The reduced major axis regression
method minimizes the sum of the areas of rectangles defined between the observed data points and the
nearest point on the line in the scatter diagram to obtain the estimates of regression coefficients. This is
shown in the following figure:

yi

(xi yi)

Y   0  1 X

(Xi, Yi)

xi

Reduced major axis method

The method of least absolute deviation regression considers the sum of the absolute deviation of the
observations from the line in the vertical direction in the scatter diagram as in the case of direct regression to
obtain the estimates of  0 and 1 .

No assumption is required about the form of probability distribution of  i in deriving the least squares

estimates. For the purpose of deriving the statistical inferences only, we assume that  i ' s are random

variable with E ( i )  0, Var ( i )   2 and Cov ( i ,  j )  0 for all i  j (i, j  1, 2,..., n). This assumption is

needed to find the mean, variance and other properties of the least squares estimates. The assumption that
 i ' s are normally distributed is utilized while constructing the tests of hypotheses and confidence intervals
of the parameters.

Based on these approaches, different estimates of  0 and 1 are obtained which have different statistical

properties. Among them the direct regression approach is more popular. Generally, the direct regression
estimates are referred as the least squares estimates or ordinary least squares estimates.
Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
4
Direct regression method
This method is also known as the ordinary least squares estimation. Assuming that a set of n paired
observations on ( xi , yi ), i  1, 2,..., n are available which satisfy the linear regression model y   0  1 X   .

So we can write the model for each observation as yi   0  1 xi   i , (i  1, 2,..., n) .

The direct regression approach minimizes the sum of squares

n n
S (  0 , 1 )    i2   ( yi   0  1 xi ) 2
i 1 i 1

with respect to  0 and 1 .

The partial derivatives of S (  0 , 1 ) with respect to  0 is

S (  0 , 1 ) n
 2 ( yt   0  1 xi )
 0 i 1

and the partial derivative of S (  0 , 1 ) with respect to 1 is

S (  0 , 1 ) n
 2 ( yi   0  1 xi )xi .
1 i 1

The solutions of  0 and 1 are obtained by setting

S (  0 , 1 )
0
 0
S (  0 , 1 )
 0.
1
The solutions of these two equations are called the direct regression estimators, or usually called as the
ordinary least squares (OLS) estimators of  0 and 1 .

This gives the ordinary least squares estimates b0 of  0 and b1 of 1 as

b0  y  b1 x
sxy
b1 
sxx
where
n n
1 n 1 n
sxy   ( xi  x )( yi  y ), sxx  ( xi  x ) 2 , x  i x , y   yi .
i 1 i 1 n i 1 n i 1

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
5
Further, we have
 2 S (  0 , 1 ) n

 02
  2 
i 1
(1)  2n,

 2 S (  0 , 1 ) n

12
 2 
i 1
xi2

 2 S (  0 , 1 ) n
 2 xt  2nx .
 0 1 i 1

The Hessian matrix which is the matrix of second order partial derivatives in this case is given as
  2 S (  0 , 1 )  2 S (  0 , 1 ) 
 
 02  0 1 
H*   2
  S ( ,  )  2 S (  0 , 1 ) 
 0 1

   0  1 12 
 n nx 
2  n 
 nx  xi2 
 
 i 1 
 '
 2    , x 
 x '
where   (1,1,...,1) ' is a n -vector of elements unity and x  ( x1 ,..., xn ) ' is a n -vector of observations on X .
The matrix H * is positive definite if its determinant and the element in the first row and column of H * are
positive. The determinant of H is given by
 n 
H *  2  n xi2  n 2 x 2 
 i 1 
n
 2n ( xi  x ) 2
i 1

 0.
n
The case when  (x  x )
i 1
i
2
 0 is not interesting because all the observations in this case are identical, i.e.

xi  c (some constant). In such a case there is no relationship between x and y in the context of regression
n
analysis. Since  (x  x )
i 1
i
2
 0, therefore H  0. So H is positive definite for any (  0 , 1 ); therefore

S (  0 , 1 ) has a global minimum at (b0 , b1 ).

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
6
The fitted line or the fitted linear regression model is
y  b0  b1 x.
The predicted values are
yˆi  b0  b1 xi (i  1, 2,..., n).

The difference between the observed value yi and the fitted (or predicted) value yˆi is called as a residual.

The i th residual is defined as

ei  yi ~ yˆi (i  1, 2,..., n)
 yi  yˆi
 yi  (b0  b1 xi ).

Properties of the direct regression estimators:

Unbiased property:
sxy
Note that b1  and b0  y  b1 x are the linear combinations of yi (i  1,..., n).
sxx
Therefore
n
b1   ki yi
i 1

n n
where ki  ( xi  x ) / sxx . Note that ki 1
i  0 and k x
i 1
i i  1, so

n
E (b1 )   ki E ( yi )
i 1
n
  ki (  0  1 xi ) .
i 1

 1.

This b1 is an unbiased estimator of 1 . Next

E (b0 )  E  y  b1 x 
 E   0  1 x    b1 x 
  0  1 x  1 x
 0 .

Thus b0 is an unbiased estimators of  0 .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
7
Variances:
Using the assumption that yi ' s are independently distributed, the variance of b1 is
n
Var (b1 )   ki2Var ( yi )   ki k j Cov( yi , y j )
i 1 i j i

 (x  x ) i
2

 2 i
(Cov( yi , y j )  0 as y1 ,..., yn are independent)
sxx2
 2 sxx
=
sxx2
2
= .
sxx

The variance of b0 is

Var (b0 )  Var ( y )  x 2 Var (b1 )  2 xCov( y , b1 ).

First we find that
Cov( y , b1 )  E  y  E ( y )b1  E (b1 )
 
 E  ( ci yi  1 ) 
 i 
1  
 E (  i )(  0  ci  1  ci xi   ci i )  1   i 
n  i i i i i 
1
  0  0  0  0
n
0
So
 1 x2 
Var (b0 )   2   .
 n sxx 

Covariance:
The covariance between b0 and b1 is

Cov(b0 , b1 )  Cov( y , b1 )  xVar (b1 )

x 2
  .
sxx

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
8
It can further be shown that the ordinary least squares estimators b0 and b1 possess the minimum variance

in the class of linear and unbiased estimators. So they are termed as the Best Linear Unbiased Estimators
(BLUE). Such a property is known as the Gauss-Markov theorem which is discussed later in multiple
linear regression model.

Residual sum of squares:

The residual sum of squares is given as
n
SSres   ei2
i 1
n
  ( yi  yˆi ) 2
i 1
n
  ( yi  b0  b1 xi ) 2
i 1
n
   yi  y  b1 x  b1 xi 
2

i 1
n
   ( yi  y )  b1 ( xi  x ) 
2

i 1
n n n
  ( yi  y ) 2  b12  ( xi  x ) 2  2b1  ( xi  x )( yi  y )
i 1 i 1 i 1

 s yy  b s  2b s
2 2
1 xx
2
1 xx

 s yy  b12 sxx
2
s 
 s yy   xy  sxx
 sxx 
sxy2
 s yy 
sxx
 s yy  b1sxy .
n
1 n
where s yy   ( yi  y ) 2 , y   yi .
i 1 n i 1

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
9
Estimation of  2
The estimator of  2 is obtained from the residual sum of squares as follows. Assuming that yi is normally

distributed, it follows that SSres has a  2 distribution with (n  2) degrees of freedom, so

SSres
~  2 (n  2).
2
Thus using the result about the expectation of a chi-square random variable, we have
E ( SSres )  (n  2) 2 .

Thus an unbiased estimator of  2 is

SSres
s2  .
n2
Note that SSres has only (n  2) degrees of freedom. The two degrees of freedom are lost due to estimation

of b0 and b1 . Since s 2 depends on the estimates b0 and b1 , so it is a model dependent estimate of  2 .

Estimate of variances of b0 and b1 :

The estimators of variances of b0 and b1 are obtained by replacing  2 by its estimate ˆ 2  s 2 as follows:

 (b )  s 2  1  x 
2
Var 0  
 n sxx 
and
2
 (b )  s .
Var 1
sxx
n n
It is observed that since  ( yi  yˆi )  0, so
i 1
e
i 1
i  0. In the light of this property, ei can be regarded as an

estimate of unknown  i (i  1,..., n) . This helps in verifying the different model assumptions on the basis of

the given sample ( xi , yi ), i  1, 2,..., n.

Further, note that

n
(i) xe
i 1
i i  0,

n
(ii)  yˆ e
i 1
i i  0,

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
10
 n n
(iii)  y   yˆ
i 1
i
i 1
i and

(iv) the fitted line always passes through ( x , y ).

Centered Model:
Sometimes it is useful to measure the independent variable around its mean. In such a case, model
yi   0  1 X i   i has a centered version as follows:

yi   0  1 ( xi  x )  1 x   (i  1, 2,..., n)
  0*  1 ( xi  x )   i

where  0*   0  1 x . The sum of squares due to error is given by

n n 2

S (  , 1 )       yi     1 ( xi  x )  .
*
0 i
2 *
0
i 1 i 1

Now solving
S (  0* , 1 )
0
 0*
S (  0* , 1 )
 0,
1*

we get the direct regression least squares estimates of  0* and 1 as

b0*  y
and
sxy
b1 
sxx
respectively.

Thus the form of the estimate of slope parameter 1 remains same in usual and centered model whereas the

form of the estimate of intercept term changes in the usual and centered models.

Further, the Hessian matrix of the second order partial derivatives of S (  0* , 1 ) with respect to  0* and 1

is positive definite at  0*  b0* and 1  b1 which ensures that S (  0* , 1 ) is minimized at  0*  b0* and
1  b1 .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
11
Under the assumption that E ( i )  0,Var ( i )   2 and Cov( i j )  0 for all i  j  1, 2,..., n , it follows that

E (b0* )   0* , E (b1 )  1 ,
2 2
Var (b0* )  , Var (b1 )  .
n sxx

In this case, the fitted model of yi   0*  1 ( xi  x )   i is

y  y  b1 ( x  x ),

and the predicted values are

yˆi  y  b1 ( xi  x ) (i  1,..., n).
Note that in centered model
Cov(b0* , b1 )  0.

No intercept term model:

Sometimes in practice, a model without an intercept term is used in those situations when xi  0  yi  0 for

all i  1, 2,..., n . A no-intercept model is

yi  1 xi   i (i  1, 2,.., n).

For example, in analyzing the relationship between the velocity ( y ) of a car and its acceleration ( X ) , the
velocity is zero when acceleration is zero.

Using the data ( xi , yi ), i  1, 2,..., n, the direct regression least squares estimate of 1 is obtained by
n n
minimizing S ( 1 )    i2   ( yi  1 xi ) 2 and solving
i 1 i 1

S ( 1 )
0
1

gives the estimator of 1 as

n

yx i i
b 
*
1
i 1
n
.
 xi2
i 1

The second order partial derivative of S ( 1 ) with respect to 1 at 1  b1 is positive which insures that b1

minimizes S ( 1 ).

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
12
Using the assumption that E ( i )  0,Var ( i )   2 and Cov( i j )  0 for all i  j  1, 2,..., n , the properties

of b1* can be derived as follows:

n

 x E( y ) i i
E (b ) *
1
i 1
n

x
i 1
2
i

x  2
i 1
 i 1
n

x i 1
2
i

 1

This b1* is an unbiased estimator of 1 . The variance of b1* is obtained as follows:

n

 x Var ( y ) 2
i i
Var (b1* )  i 1
2
 n 2
  xi 
 i 1 
n

x 2
i
2 i 1
2
 n 2
  xi 
 i 1 
2
 n

xi 1
2
i

and an unbiased estimator of  2 is obtained as

n n

y
i 1
2
i  b1  yi xi
i 1
.
n 1

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
13
Maximum likelihood estimation
We assume that  i ' s (i  1, 2,..., n) are independent and identically distributed following a normal

distribution N (0,  2 ). Now we use the method of maximum likelihood to estimate the parameters of the
linear regression model
yi   0  1 xi   i (i  1, 2,..., n),

the observations yi (i  1, 2,..., n) are independently distributed with N (  0  1 xi ,  2 ) for all i  1, 2,..., n.

The likelihood function of the given observations ( xi , yi ) and unknown parameters  0 , 1 and  2 is
1/ 2
n
 1   1 
L( xi , yi ;  0 , 1 ,  2 )    2 
exp   2 ( yi   0  1 xi ) 2 .
i 1  2   2 
The maximum likelihood estimates of  0 , 1 and  2 can be obtained by maximizing L( xi , yi ;  0 , 1 ,  2 ) or

equivalently in ln L( xi , yi ;  0 , 1 ,  2 ) where

n n  1  n
ln L( xi , yi ;  0 , 1 ,  2 )     ln 2    ln  2   2   ( yi   0  1 xi ) 2 .
2 2  2  i 1
The normal equations are obtained by partial differentiation of log-likelihood with respect to  0 , 1 and  2
and equating them to zero as follows:
 ln L( xi , yi ;  0 , 1 ,  2 ) 1 n

 0
 2

(y  
i 1
i 0  1 xi )  0

 ln L( xi , yi ;  0 , 1 ,  2 ) 1 n

1
 2

(y  
i 1
i 0  1 xi )xi  0

and
 ln L( xi , yi ;  0 , 1 ,  2 ) n 1 n

 2
 2  4
2 2
(y  
i 1
i 0  1 xi ) 2  0.

The solution of these normal equations give the maximum likelihood estimates of  0 , 1 and  2 as

b0  y  b1 x
n

 ( x  x )( y  y )
i i
sxy
b1  i 1
n

 (x  x ) 2 sxx
i
i 1

and
n

 ( y  b i 0  b1 xi ) 2
s 2  i 1

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
14
respectively.

It can be verified that the Hessian matrix of second order partial derivation of ln L with respect to  0 , 1 ,

and  2 is negative definite at  0  b0 , 1  b1 , and  2  s 2 which ensures that the likelihood function is
maximized at these values.

Note that the least squares and maximum likelihood estimates of  0 and 1 are identical. The least squares

and maximum likelihood estimates of  2 are different. In fact, the least squares estimate of  2 is
1 n
s2   ( yi  y )2
n  2 i 1
so that it is related to maximum likelihood estimate as
n2 2
s 2  s .
n

Thus b0 and b1 are unbiased estimators of  0 and 1 whereas s 2 is a biased estimate of  2 , but it is

asymptotically unbiased. The variances of b0 and b1 are same as of b0 and b1 respectively but

Var ( s 2 )  Var ( s 2 ).

Testing of hypotheses and confidence interval estimation for slope parameter:

Now we consider the tests of hypothesis and confidence interval estimation for the slope parameter of the
model under two cases, viz., when  2 is known and when  2 is unknown.

Case 1: When  2 is known:

Consider the simple linear regression model yi   0  1 xi   i (i  1, 2,..., n) . It is assumed that  i ' s are

independent and identically distributed and follow N (0,  2 ).

First we develop a test for the null hypothesis related to the slope parameter
H 0 : 1  10

where 10 is some given constant.

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
15
 2
Assuming  2 to be known, we know that E (b1 )  1 , Var (b1 )  and b1 is a linear combination of
sxx

normally distributed yi ' s . So

 2 
b1 ~ N  1 , 
 sxx 

and so the following statistic can be constructed

b1  10
Z1 
2
sxx

which is distributed as N (0,1) when H 0 is true.

A decision rule to test H1 : 1  10 can be framed as follows:

Reject H 0 if Z1  Z / 2

where Z /2 is the  / 2 percent points on normal distribution.

Similarly, the decision rule for one sided alternative hypothesis can also be framed.

The 100 (1   )% confidence interval for 1 can be obtained using the Z1 statistic as follows:

P   z /2  Z1  z /2   1  
 
 
 b1  1
P  z /2   z /2   1  
  2 
 
 sxx 
 2 2 
P b1  z /2  1  b1  z /2   1.
 sxx sxx 

So 100 (1   )% confidence interval for 1 is

 2 2 
b1  z / 2 , b1  z / 2 
 sxx sxx 

where z / 2 is the  / 2 percentage point of the N (0,1) distribution.

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
16
Case 2: When  2 is unknown:
When  2 is unknown then we proceed as follows. We know that
SSres
~  2 (n  2)
2
and
 SS 
E  res    2 .
 n2
Further, SSres /  2 and b1 are independently distributed. This result will be proved formally later in next
module on multiple linear regression. This result also follows from the result that under normal distribution,
the maximum likelihood estimates, viz., the sample mean (estimator of population mean) and the sample
variance (estimator of population variance) are independently distributed, so b1 and s 2 are also
independently distributed.
Thus the following statistic can be constructed:
b1  1
t0 
ˆ 2
sxx
b1  1

SSres
(n  2) sxx

which follows a t -distribution with (n  2) degrees of freedom, denoted as tn  2 , when H 0 is true.

A decision rule to test H1 : 1  10 is to

reject H 0 if t0  tn  2, / 2

where tn  2, / 2 is the  / 2 percent point of the t -distribution with (n  2) degrees of freedom. Similarly, the

decision rule for one sided alternative hypothesis can also be framed.

The 100 (1   )% confidence interval of 1 can be obtained using the t0 statistic as follows:

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
17
Consider
P  t /2  t0  t /2   1  
 
 
 b1  1
P t /2   t /2   1  
 ˆ 2 
 
 sxx 
 ˆ 2 ˆ 2 
P b1  t /2  1  b1  t / 2   1.
 sxx sxx 

So the 100 (1   )% confidence interval 1 is

 SSres SS res 
b1  tn  2, /2 , b1  tn  2, /2 .
 (n  2) sxx (n  2) sxx 

Testing of hypotheses and confidence interval estimation for intercept term:

Now, we consider the tests of hypothesis and confidence interval estimation for intercept term under two
cases, viz., when  2 is known and when  2 is unknown.

Case 1: When  2 is known:

Suppose the null hypothesis under consideration is
H 0 :  0   00 ,

 1 x2 
where  2 is known, then using the result that E (b0 )   0 , Var (b0 )   2    and b0 is a linear
 n sx 
combination of normally distributed random variables, the following statistic
b0   00
Z0 
 1 x2 
   
2

 n sxx 

has a N (0,1) distribution when H 0 is true.

A decision rule to test H1 :  0   00 can be framed as follows:

Reject H 0 if Z 0  Z /2

where Z /2 is the  / 2 percentage points on normal distribution. Similarly, the decision rule for one sided
alternative hypothesis can also be framed.
Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
18
The 100 (1   )% confidence intervals for  0 when  2 is known can be derived using the Z 0 statistic as
follows:
P   z /2  Z 0  z /2   1  
 
 
 b0   0 
P   z /2   z /2   1  
  1 x2  
2  
  n s xx 

 
 
21 x2  21 x2 
P b0  z /2   
    0  b0  z /2       1.
  n sxx   n sxx  

So the 100 (1   )% of confidential interval of  0 is

  1 x2  
21 x2
b0  z / 2  2   , 
 0  /2
b z    .
  n sxx   n sxx  

Case 2: When  2 is unknown:

When  2 is unknown, then the following statistic is constructed
b0   00
t0 
SSres  1 x 2 
  
n  2  n sxx 

which follows a t -distribution with (n  2) degrees of freedom, i.e., tn  2 when H 0 is true.

A decision rule to test H1 :  0   00 is as follows:

Reject H 0 whenever t0  tn  2, / 2

where tn  2, / 2 is the  / 2 percentage point of the t -distribution with (n  2) degrees of freedom. Similarly,

the decision rule for one sided alternative hypothesis can also be framed.

The 100 (1   )% confidence interval of  0 can be obtained as follows:

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
19
Consider
P tn  2, /2  t0  tn  2, /2   1  
 
 
 b0   0 
P tn  2, /2   tn  2, /2   1  
 SS res  1 x 2  
    
 n  2  n sxx  
 SSres  1 x 2  SS res  1 x 2 
P b0  tn  2, /2      0  b0  t n  2, /2      1.
 n  2  n sxx  n  2  n sxx  

So 100(1   )% confidence interval for  0 is

 SS res  1 x 2  SSres  1 x 2 
b0  tn  2, / 2    , b0  tn  2, / 2   .
 n  2  n sxx  n  2  n sxx  

Test of hypothesis for  2

We have considered two types of test statistics for testing the hypothesis about the intercept term and slope
parameter- when  2 is known and when  2 is unknown. While dealing with the case of known  2 , the
value of  2 is known from some external sources like past experience, long association of the experimenter
with the experiment, past studies etc. In such situations, the experimenter would like to test the hypothesis
like H 0 :  2   02 against H 0 :  2   02 where  02 is specified. The test statistic is based on the result
SS r es
~  n2 2 . So the test statistic is
2
SS r es
C0  ~  n2 2 under H 0 .
 02
The decision rule is to reject H 0 if C0   n2 2, /2 or C0   n2 2,1 /2 .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
20
Confidence interval for  2
A confidence interval for  2 can also be derived as follows. Since SSres /  2 ~  n2 2 , thus consider

 SS 
P   n2 2, /2  res   n2 2,1 /2   1  
  2

 SS SS 
P  2 res   2  2 res   1   .
  n  2,1 / 2  n 2, / 2 

The corresponding 100(1   )% confidence interval for  2 is

 SSres SS 
 2 , 2 res  .
  n  2,1 / 2  n  2, / 2 

Joint confidence region for  0 and 1 :

A joint confidence region for  0 and 1 can also be found. Such region will provide a 100(1   )%

confidence that both the estimates of  0 and 1 are correct. Consider the centered version of the linear
regression model
yi   0*  1 ( xi  x )   i

where  0*   0  1 x . The least squares estimators of  0* and 1 are

sxy
b0*  y and b1  ,
sxx
respectively.

Using the results that

E (b0* )   0* ,
E (b1 )  1 ,
2
Var (b0* )  ,
n
2
Var (b1 )  .
sxx

When  2 is known, then the statistic

b0*   0* b1  1
~ N (0,1) and ~ N (0,1).
 2
2
n sxx
Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
21
Moreover, both the statistics are independently distributed. Thus
2
 
2
 
 * * 
 
 b0   0  ~ 12 and  b1  1  ~  2
 2   2  1

  
   s 
 n   xx 
are also independently distributed because b0* and b1 are independently distributed. Consequently, the sum
of these two
n(b0*   o* ) 2 sxx (b1  1 ) 2
 ~  22 .
 2
 2

Since
SSres
~  n2 2
2
and SSres is independently distributed of b0* and b1 , so the ratio

 n(b0*   0* ) 2 sxx (b1  1 ) 2 

   2
 2 2  ~ F2, n  2 .
 res 
SS
 2  (n  2)
  
Substituting b0*  b0  b1 x and  0*   0  1 x , we get

 n  2   Qf 
  
 2   SSres 
where
n n
Q f  n(b0   0 ) 2  2 xt (b0  1 )(b1  1 )   xi2 (b1  1 ) 2 .
i 1 i 1

Since
 n  2  Q f 
P    F2, n  2   1  
 2  SSres 
holds true for all values of  0 and 1 , so the 100 (1   ) % confidence region for  0 and 1 is

 n  2  Qf
 .  F2, n  2;1 . .
 2  SSres
This confidence region is an ellipse which gives the 100 (1   )% probability that  0 and 1 are contained

simultaneously in this ellipse.

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
22
Analysis of variance:
The technique of analysis of variance is usually used for testing the hypothesis related to equality of more
than one parameters, like population means or slope parameters. It is more meaningful in case of multiple
regression model when there are more than one slope parameters. This technique is discussed and
illustrated here to understand the related basic concepts and fundamentals which will be used in developing
the analysis of variance in the next module in multiple linear regression model where the explanatory
variables are more than two.

A test statistic for testing H 0 : 1  0 can also be formulated using the analysis of variance technique as

follows.

On the basis of the identity

yi  yˆi  ( yi  y )  ( yˆi  y ),
the sum of squared residuals is
n
S (b)   ( yi  yˆi ) 2
i 1
n n n
  ( yi  y ) 2   ( yˆi  yi ) 2  2 ( yi  y )( yˆi  y ).
i 1 i 1 i 1

Further consider
n n

 ( yi  y )( yˆi  y )   ( yi  y )b1 ( xi  x )
i 1 i 1
n
 b12  ( xi  x ) 2
i 1
n
  ( yˆi  y ) 2 .
i 1

Thus we have
n n n

 i 1
( yi  y ) 2   ( yi  yˆi ) 2   ( yˆi  y ) 2 .
i 1 i 1

n
The term  ( y  y)
i 1
i
2
is called the sum of squares about the mean, corrected sum of squares of y (i.e.,

SScorrected), total sum of squares, or s yy .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
23
 n
The term  ( y  yˆ )
i 1
i i
2
describes the deviation: observation minus predicted value, viz., the residual sum of

n
squares, i.e., SS res   ( yi  yˆi ) 2
i 1

n
whereas the term  ( yˆ  y )
i 1
i
2
describes the proportion of variability explained by regression,

n
SS r e g   ( yˆi  y ) 2 .
i 1

n
If all observations yi are located on a straight line, then in this case  ( y  yˆ )
i 1
i i
2
 0 and thus

SScorrected  SS r e g .

Note that SSr e g is completely determined by b1 and so has only one degrees of freedom. The total sum of
n n
squares s yy   ( yi  y ) 2 has (n  1) degrees of freedom due to constraint  ( y  y)  0
i and SS res has
i 1 i 1

(n  2) degrees of freedom as it depends on the determination of b0 and b1 .

All sums of squares are mutually independent and distributed as  df2 with df degrees of freedom if the

errors are normally distributed.

The mean square due to regression is

SS r e g
MS r e g 
1
and mean square due to residuals is
SSres
MSE  .
n2
The test statistic for testing H 0 : 1  0 is

MS r e g
F0  .
MSE
If H 0 : 1  0 is true, then MSr e g and MSE are independently distributed and thus

F0 ~ F1, n  2 .
Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
24
The decision rule for H1 : 1  0 is to reject H 0 if

F0  F1,n  2;1

at  level of significance. The test procedure can be described in an Analysis of variance table.

Analysis of variance for testing H 0 : 1  0

Source of variation Sum of squares Degrees of freedom Mean square F

Regression SSr e g 1 MSr e g MSr e g / MSE

Residual SS res n2 MSE

Total s yy n 1

Some other forms of SSreg , SSres and s yy can be derived as follows:

The sample correlation coefficient then may be written as

sxy
rxy  .
sxx s yy

Moreover, we have
sxy s yy
b1   rxy .
sxx sxx

The estimator of  2 in this case may be expressed as

1 n 2
s2   ei
n  2 i 1
1
 SS res .
n2
Various alternative formulations for SS res are in use as well:
n
SSres   [ yi  (b0  b1 xi )]2
i 1
n
  [( yi  y )  b1 ( xi  x )]2
i 1

 s yy  b12 sxx  2b1sxy

 s yy  b12 sxx
( sxy ) 2
 s yy  .
sxx

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
25
Using this result, we find that
SScorrected  s yy

and
SSr e g  s yy  SS res
( sxy ) 2

sxx
b s 2
1 xx

 b1sxy .

Goodness of fit of regression

It can be noted that a fitted model can be said to be good when residuals are small. Since SSres is based on

residuals, so a measure of quality of fitted model can be based on SSres . When intercept term is present in the
model, a measure of goodness of fit of the model is given by
SS res
R2  1 
s yy
SS r e g
 .
s yy

This is known as the coefficient of determination. This measure is based on the concept that how much
variation in y ’s stated by s yy is explainable by SSreg and how much unexplainable part is contained in

SS res . The ratio SS r e g / s yy describes the proportion of variability that is explained by regression in relation

to the total variability of y . The ratio SSres / s yy describes the proportion of variability that is not covered

by the regression.

It can be seen that

R 2  rxy2

where rxy is the simple correlation coefficient between x and y. Clearly 0  R 2  1 , so a value of R 2 closer

to one indicates the better fit and value of R 2 closer to zero indicates the poor fit.

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
26
Prediction of values of study variable
An important use of linear regression modeling is to predict the average and actual values of study variable.
The term prediction of value of study variable corresponds to knowing the value of E ( y ) (in case of
average value) and value of y (in case of actual value) for a given value of explanatory variable. We
consider both the cases.

Case 1: Prediction of average value

Under the linear regression model y   0  1 x   , the fitted model is y  b0  b1 x where b0 and b1 are the

OLS estimators of  0 and 1 respectively.

Suppose we want to predict the value of E ( y ) for a given value of x  x0 . Then the predictor is given by

E ( y | x0 )  ˆ y / x0  b0  b1 x0 .

Predictive bias
Then the prediction error is given as
ˆ y| x  E ( y )  b0  b1 x0  E (  0  1 x0   )
0

 b0  b1 x0  (  0  1 x0 )
 (b0   0 )  (b1  1 ) x0 .
Then
E  ˆ y| x0  E ( y )   E (b0   0 )  E (b1  1 ) x0
 00  0
Thus the predictor  y / x0 is an unbiased predictor of E ( y ).

Predictive variance:
The predictive variance of ˆ y| x0 is

PV ( ˆ y| x0 )  Var (b0  b1 x0 )
 Var  y  b1 ( x0  x ) 
 Var ( y )  ( x0  x ) 2 Var (b1 )  2( x0  x )Cov( y , b1 )
2  2 ( x0  x ) 2
  0
n sxx
 1 ( x  x )2 
2   0 .
n sxx 

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
27
Estimate of predictive variance
The predictive variance can be estimated by substituting  2 by ˆ 2  MSE as

 ( ˆ )  ˆ 2  1  ( x0  x ) 
2
PV y| x0  
n sxx 
 1 ( x0  x ) 2 
 MSE   .
n sxx 

Prediction interval estimation:

The 100(1-  )% prediction interval for E ( y / x0 ) is obtained as follows:

The predictor ˆ y| x0 is a linear combination of normally distributed random variables, so it is also normally

distributed as


ˆ y| x ~ N  0  1 x0 , PV  ˆ y| x
0 0
 .
So if  2 is known, then the distribution of
ˆ y| x  E ( y | x0 )
0

PV ( ˆ y| x0 )

is N (0,1). So the 100(1-  )% prediction interval is obtained as

 ˆ y| x0  E ( y | x0 ) 
P   z /2   z /2   1  
 PV ( ˆ y| x0 ) 

which gives the prediction interval for E ( y / x0 ) as

  1 ( x  x )2  ( x0  x ) 2  
2 1
 ˆ y| x0  z /2  2   0 , 
ˆ 
 y| x0  /2
z     .
 n sxx  n sxx  


When  2 is unknown, it is replaced by ˆ 2  MSE and in this case the sampling distribution of
ˆ y|x  E ( y | x0 )
0

 1 ( x  x )2 
MSE   0
n sxx 

is t -distribution with (n  2) degrees of freedom, i.e., tn  2 .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
28
The 100(1-  )% prediction interval in this case is
 
 
 ˆ y| x0  E ( y | x0 ) 
P  t /2,n  2   t   1   .
  1 ( x  x )2  2
,n2

MSE   0 
  n s  
 xx 
which gives the prediction interval as
  1 ( x  x )2   1 ( x0  x ) 2  
 ˆ y| x0  t /2, n  2 MSE   0  , 
ˆ y| x0  t /2, n  2 MSE    .
 n sxx  n sxx  

Note that the width of prediction interval E ( y | x0 ) is a function of x0 . The interval width is minimum for

x0  x and widens as x0  x increases. This is expected also as the best estimates of y to be made at x -

values lie near the center of the data and the precision of estimation to deteriorate as we move to the
boundary of the x -space.

Case 2: Prediction of actual value

If x0 is the value of the explanatory variable, then the actual value predictor for y is

ŷ0  b0  b1 x0 .

The true value of y in the prediction period is given by y0   0  1 x0   0 where  0 indicates the value that
would be drawn from the distribution of random error in the prediction period. Note that the form of
predictor is the same as of average value predictor but its predictive error and other properties are different.
This is the dual nature of predictor.

Predictive bias:
The predictive error of ŷ0 is given by

yˆ 0  y0  b0  b1 x0  (  0  1 x0   0 )
 (b0   0 )  (b1  1 ) x0   .
Thus, we find that
E ( yˆ 0  y0 )  E (b0   0 )  E (b1  1 ) x0  E ( 0 )
 000  0

which implies that ŷ0 is an unbiased predictor of y0 .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
29
Predictive variance
Because the future observation y0 is independent of ŷ0 , the predictive variance of ŷ0 is

PV ( yˆ 0 )  E ( yˆ 0  y0 ) 2
 E[(b0   0 )  ( x0  x )(b1  1 )  (b1  1 ) x   0 ]2
 Var (b0 )  ( x0  x ) 2 Var (b1 )  x 2Var (b1 )  Var ( 0 )  2( x0  x )Cov(b0 , b1 )  2 xCov(b0 , b1 )  2( x0  x )Var (b1 )
[rest of the terms are 0 assuming the independence of  0 with 1 ,  2 ,...,  n ]
 Var (b0 )  [( x0  x ) 2  x 2  2( x0  x )]Var (b1 )  Var ( )  2[( x0  x )  2 x ]Cov(b0 , b1 )
 Var (b0 )  x02Var (b1 )  Var ( 0 )  2 x0Cov(b0 , b1 )
1 x2  2 x 2
  2     x02   2  2 x0
 n sxx  sxx sxx
 1 ( x  x )2 
  2 1   0 .
 n sxx 

Estimate of predictive variance

The estimate of predictive variance can be obtained by replacing  2 by its estimate ˆ 2  MSE as
  1 ( x  x )2 
PV ( yˆ 0 )  ˆ 2 1   0 
 n sxx 
 1 ( x  x )2 
 MSE 1   0 .
 n sxx 

Prediction interval:
If  2 is known, then the distribution of
yˆ 0  y0
PV ( yˆ 0 )

is N (0,1). So the 100(1-  )% prediction interval is obtained as

 yˆ  y0 
P   z /2  0  z /2   1  
 PV ( yˆ 0 ) 
which gives the prediction interval for y0 as

  1 ( x  x )2  1 ( x0  x ) 2  
2
 yˆ 0  z /2  2 1   0 , ˆ 
 0  /2
y z   1   .
  n sxx   n sxx  

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
30
When  2 is unknown, then
yˆ 0  y0
 ( yˆ )
PV 0

follows a t -distribution with (n  2) degrees of freedom. The 100(1-  )% prediction interval for ŷ0 in this

case is obtained as
 yˆ  y0 
P  t /2,n  2  0  t /2, n  2   1  
  ( yˆ )
PV 
 0 
which gives the prediction interval
  1 ( x  x )2   1 ( x0  x ) 2  
 yˆ 0  t /2,n  2 MSE 1   0 , ˆ 
 0  /2,n  2
y t MSE 1    .
  n s xx   n s xx  

The prediction interval is of minimum width at x0  x and widens as x0  x increases.

The prediction interval for ŷ0 is wider than the prediction interval for ˆ y / x0 because the prediction interval

for ŷ0 depends on both the error from the fitted model as well as the error associated with the future
observations.

Reverse regression method

The reverse (or inverse) regression approach minimizes the sum of squares of horizontal distances between
the observed data points and the line in the following scatter diagram to obtain the estimates of regression
parameters.
yi

Y   0  1 X
(xi,
(Xi,
)

x,
Reverse regression
Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
31
The reverse regression has been advocated in the analysis of gender (or race) discrimination in salaries. For
example, if y denotes salary and x denotes qualifications and we are interested in determining if there is a
gender discrimination in salaries, we can ask:
“Whether men and women with the same qualifications (value of x) are getting the same salaries
(value of y). This question is answered by the direct regression.”

Alternatively, we can ask:

“Whether men and women with the same salaries (value of y) have the same qualifications (value of
x). This question is answered by the reverse regression, i.e., regression of x on y.”

The regression equation in case of reverse regression can be written as

xi   0*  1* yi   i (i  1, 2,..., n)

where  i ’s are the associated random error components and satisfy the assumptions as in the case of usual

simple linear regression model. The reverse regression estimates ˆOR of  0* and ˆ1R of  1* for the model

are obtained by interchanging the x and y in the direct regression estimators of  0 and 1 . The estimates are
obtained as
ˆOR  x  ˆ1R y
and
s yy
ˆ1R 
sxy

for  0 and 1 respectively. The residual sum of squares in this case is

sxy2
*
SS res  sxx  .
s yy

Note that
sxy2
ˆ1R b1   rxy2
sxx s yy

where b1 is the direct regression estimator of slope parameter and rxy is the correlation coefficient between x

and y. Hence if rxy2 is close to 1, the two regression lines will be close to each other.

An important application of reverse regression method is in solving the calibration problem.

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
32
Orthogonal regression method (or major axis regression method)
The direct and reverse regression methods of estimation assume that the errors in the observations are either
in x -direction or y -direction. In other words, the errors can be either in dependent variable or independent
variable. There can be situations when uncertainties are involved in dependent and independent variables
both. In such situations, the orthogonal regression is more appropriate. In order to take care of errors in both
the directions, the least squares principle in orthogonal regression minimizes the squared perpendicular
distance between the observed data points and the line in the following scatter diagram to obtain the
estimates of regression coefficients. This is also known as major axis regression method. The estimates
obtained are called as orthogonal regression estimates or major axis regression estimates of regression
coefficients.

yi

(xi, yi)

Y   0  1 X

(Xi, Yi)

xi

Orthogonal or major axis regression

If we assume that the regression line to be fitted is Yi   0  1 X i , then it is expected that all the

observations ( xi , yi ), i  1, 2,..., n lie on this line. But these points deviate from the line and in such a case,

the squared perpendicular distance of observed data ( xi , yi ) (i  1, 2,..., n) from the line is given by

di2  ( X i  xi ) 2  (Yi  yi ) 2

where ( X i , Yi ) denotes the i th pair of observation without any error which lie on the line.

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
33
 n
The objective is to minimize the sum of squared perpendicular distances given by i 1
di2 to obtain the

estimates of  0 and 1 . The observations ( xi , yi ) (i  1, 2,..., n) are expected to lie on the line

Yi   0  1 X i ,
so let
Ei  Yi   0  1 X i  0.

n
The regression coefficients are obtained by minimizing d
i 1
i
2
under the constraints Ei ' s using the

Lagrangian’s multiplier method. The Lagrangian function is

n n
L0   di2  2 i Ei
i 1 i 1

where 1 ,..., n are the Lagrangian multipliers. The set of equations are obtained by setting

L0 L L L
 0, 0  0, 0  0 and 0  0 (i  1, 2,..., n).
X i Yi  0 1
Thus we find
L0
 ( X i  xi )  i 1  0
X i
L0
 (Yi  yi )  i  0
Yi
L0 n

 0
 
i 1
i 0

L0 n

1
  X
i 1
i i  0.

Since
X i  xi  i 1
Yi  yi  i ,

so substituting these values is  i , we obtain

Ei  ( yi  i )   0  1 ( xi  i 1 )  0
 0  1 xi  yi
 i  .
1  12
n
Also using this i in the equation 
i 1
i  0 , we get

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
34
 n

 ( 0  1 xi  yi )
i 1
0
1  12
n
and using ( X i  xi )  i 1  0 and  X
i 1
i i  0 , we get

  ( x    )  0.
i 1
i i i 1

Substituting i in this equation, we get

n

 ( x   x 0 i
2
1 i  yi xi )
1 (  0  1 xi  yi ) 2
i 1
  0. (1)
(1  i2 ) (1  12 ) 2
n
Using i in the equation and using the equation 
i 1
i  0 , we solve

 ( 0  1 xi  yi )
i 1
 0.
1  12

The solution provides an orthogonal regression estimate of  0 as

ˆ0OR  y  ˆ1OR x

where ˆ1OR is an orthogonal regression estimate of 1.

Now, substituting  0OR in equation (1), we get

2

)  yxi  1 xxi   x  xi yi   1   y  1 x  1 xi  yi   0
n n

 (1  
i 1
1
2 2
1 i
i 1

or
n n 2

(1  12 )  xi  yi  y  1 ( xi  x )  1   ( yi  y )  1 ( xi  x )  0
i 1 i 1

or
n n
(1  12 ) (ui  x )(vi  1ui )  1  (vi  1ui ) 2  0
i 1 i 1

where ui  xi  x ,
vi  yi  y .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
35
 n n
Since u  u
i 1
i
i 1
i  0, so
n

  
i 1
u v  1 (ui2  vi2 )  ui vi   0
2
1 i i

or
12 sxy  1 ( sxx  s yy )  sxy  0.

Solving this quadratic equation provides the orthogonal regression estimate of 1 as

s  sxx   sign  sxy  ( sxx  s yy ) 2  4s 2xy

ˆ1OR 
yy

2sxy

where sign( sxy ) denotes the sign of sxy which can be positive or negative . So

1 if sxy  0
sign( sxy )   .
1 if sxy  0.

n
Notice that this gives two solutions for ˆ1OR . We choose the solution which minimizes d
i 1
i
2
. The other

n
solution maximizes d
i 1
i
2
and is in the direction perpendicular to the optimal solution. The optimal

solution can be chosen with the sign of sxy .

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
36
Reduced major axis regression method:
The direct, reverse and orthogonal methods of estimation minimize the errors in a particular direction which
is usually the distance between the observed data points and the line in the scatter diagram. Alternatively,
one can consider the area extended by the data points in certain neighbourhood and instead of distances, the
area of rectangles defined between corresponding observed data point and nearest point on the line in the
following scatter diagram can also be minimized. Such an approach is more appropriate when the
uncertainties are present in study and explanatory variables both. This approach is termed as reduced major
axis regression.
yi

(xi yi)

Y   0  1 X

(Xi, Yi)

xi

Reduced major axis method

Suppose the regression line is Yi   0  1 X i on which all the observed points are expected to lie. Suppose

the points ( xi , yi ), i  1, 2,..., n are observed which lie away from the line. The area of rectangle extended

between the i th observed data point and the line is

Ai  ( X i ~ xi )(Yi ~ yi ) (i  1, 2,..., n)

where ( X i , Yi ) denotes the i th pair of observation without any error which lie on the line.

The total area extended by n data points is

n n

 A  (X
i 1
i
i 1
i ~ xi )(Yi ~ yi ).

All observed data points ( xi , yi ), (i  1, 2,..., n) are expected to lie on the line

Yi   0  1 X i

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
37
and let
Ei*  Yi   0  1 X i  0.

So now the objective is to minimize the sum of areas under the constraints Ei* to obtain the reduced major
axis estimates of regression coefficients. Using the Lagrangian multiplies method, the Lagrangian function is
n n
LR   Ai   i Ei*
i 1 i 1
n n
  ( X i  xi )(Yi  yi )   i Ei*
i 1 i 1

where 1 ,..., n are the Lagrangian multipliers. The set of equations are obtained by setting

LR L L L
 0, R  0, R  0, R  0 (i  1, 2,..., n).
X i Yi  0 1
Thus
LR
 (Yi  yi )  1i  0
X i
LR
 ( X i  xi )  i  0
Yi
LR n
  i  0
 0 i 1
LR n
  i X i  0.
1 i 1
Now
X i  xi  i
Yi  yi  1i
 0  1 X i  yi  1i
 0  1 ( xi  i )  yi  1i
y   0  1 xi
 i  i .
2 1
n
Substituting i in 
i 1
i 0, the reduced major axis regression estimate of  0 is obtained as

ˆ0 RM  y  ˆ1RM x

where ˆ1RM is the reduced major axis regression estimate of 1 . Using X i  xi  i , i and ˆ0 RM in
n

 X
i 1
i i  0 , we get

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
38
 n
 yi  y  1 x  1 xi  yi  y  1 x  1 xi 

i 1  21
 xi 
21
  0.
 
Let ui  xi  x and vi  yi  y , then this equation can be re-expressed as
n

 (v   u )(v   u  2 x )  0.
i 1
i 1 i i 1 i 1

n n
Using u  u
i 1
i
i 1
i  0, we get

n n

v
i 1
2
i  12  ui2  0.
i 1

Solving this equation, the reduced major axis regression estimate of 1 is obtained as

s yy
ˆ1RM  sign( sxy )
sxx

1 if sxy  0
where sign ( sxy )  
1 if sxy  0.
We choose the regression estimator which has same sign as of sxy .

Least absolute deviation regression method

The least squares principle advocates the minimization of sum of squared errors. The idea of squaring the
errors is useful in place of simple errors because the random errors can be positive as well as negative. So
consequently their sum can be close to zero indicating that there is no error in the model and which can be
misleading. Instead of the sum of random errors, the sum of absolute random errors can be considered
which avoids the problem due to positive and negative random errors.

In the method of least squares, the estimates of the parameters  0 and 1 in the model
n
yi   0  1 xi   i . (i  1, 2,..., n) are chosen such that the sum of squares of deviations 
i 1
i
2
is minimum. In

the method of least absolute deviation (LAD) regression, the parameters  0 and 1 are estimated such that
n
the sum of absolute deviations 
i 1
i is minimum. It minimizes the absolute vertical sum of errors as in the

following scatter diagram:

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
39
 yi

(xi, yi)

Y   0  1 X

(Xi, Yi)

xi
Least absolute deviation regression

The LAD estimates ˆ0 L and ˆ1L are the estimates of  0 and 1 , respectively which minimize
n
LAD(  0 , 1 )   yi   0  1 xi
i 1

for the given observations ( xi , yi ) (i  1, 2,..., n).

Conceptually, LAD procedure is simpler than OLS procedure because e (absolute residuals) is a more

straightforward measure of the size of the residual than e 2 (squared residuals). The LAD regression
estimates of  0 and 1 are not available in closed form. Rather they can be obtained numerically based on
algorithms. Moreover, this creates the problems of non-uniqueness and degeneracy in the estimates. The
concept of non-uniqueness relates to that more than one best lines pass through a data point. The degeneracy
concept describes that the best line through a data point also passes through more than one other data points.
The non-uniqueness and degeneracy concepts are used in algorithms to judge the quality of the estimates.
The algorithm for finding the estimators generally proceeds in steps. At each step, the best line is found that
passes through a given data point. The best line always passes through another data point, and this data point
is used in the next step. When there is non-uniqueness, then there are more than one best lines. When there
is degeneracy, then the best line passes through more than one other data point. When either of the problem
is present, then there is more than one choice for the data point to be used in the next step and the algorithm
may go around in circles or make a wrong choice of the LAD regression line. The exact tests of hypothesis
and confidence intervals for the LAD regression estimates can not be derived analytically. Instead they are
derived analogous to the tests of hypothesis and confidence intervals related to ordinary least squares
estimates.
Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
40
Estimation of parameters when X is stochastic
In a usual linear regression model, the study variable is supped to be random and explanatory variables are
assumed to be fixed. In practice, there may be situations in which the explanatory variable also becomes
random.

Suppose both dependent and independent variables are stochastic in the simple linear regression model
y   0  1 X  

where  is the associated random error component. The observations ( xi , yi ), i  1, 2,..., n are assumed to be
jointly distributed. Then the statistical inferences can be drawn in such cases which are conditional on X .

Assume the joint distribution of X and y to be bivariate normal N (  x ,  y ,  x2 ,  y2 ,  ) where  x and  y

are the means of X and y;  x2 and  y2 are the variances of X and y; and  is the correlation coefficient

between X and y . Then the conditional distribution of y given X  x is univariate normal conditional
mean
E ( y | X  x)   y| x   0  1 x

and conditional variance of y given X  x is

Var ( y | X  x)   y2| x   y2 (1   2 )

where
 0   y   x 1
and
y
1  .
x
When both X and y are stochastic, then the problem of estimation of parameters can be reformulated as
follows. Consider a conditional random variable y | X  x having a normal distribution with mean as

conditional mean  y| x and variance as conditional variance Var ( y | X  x)   y2| x . Obtain n independently

distributed observation yi | xi , i  1, 2,..., n from N (  y| x ,  y2| x ) with nonstochastic X . Now the method of

maximum likelihood can be used to estimate the parameters which yields the estimates of  0 and 1 as
earlier in the case of nonstochastic X as
b  y  b1 x

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
41
and
sxy
b1 
sxx
respectively.

Moreover, the correlation coefficient

E ( y   y )( X   x )

 y x
can be estimated by the sample correlation coefficient
n

 ( y  y )( x  x )
i i
ˆ  i 1
n n

 ( xi  x )2
i 1
 ( y  y)
i 1
i
2

sxy

sxx s yy
sxx
 b1 .
s yy

Thus
sxx
ˆ 2  b12
s yy
s
 b1 xy
s yy
n
s yy   ˆi2
 i 1

s yy
R 2

which is same as the coefficient of determination. Thus R 2 has the same expression as in the case when X
is fixed. Thus R 2 again measures the goodness of fitted model even when X is stochastic.

Regression Analysis | Chapter 2 | Simple Linear Regression Analysis | Shalabh, IIT Kanpur
42